SAFETY AND COMPLEXITY: INTER-DEPARTMENTAL RELATIONSHIPS School of
Inter-firm Relationships and Asset Prices · Inter-firm Relationships and Asset Prices Carlos...
Transcript of Inter-firm Relationships and Asset Prices · Inter-firm Relationships and Asset Prices Carlos...
Finance and Economics Discussion SeriesDivisions of Research & Statistics and Monetary Affairs
Federal Reserve Board, Washington, D.C.
Inter-firm Relationships and Asset Prices
Carlos Ramırez
2017-014
Please cite this paper as:Ramırez, Carlos (2017). “Inter-firm Relationships and Asset Prices,” Finance and Eco-nomics Discussion Series 2017-014. Washington: Board of Governors of the Federal ReserveSystem, https://doi.org/10.17016/FEDS.2017.014.
NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminarymaterials circulated to stimulate discussion and critical comment. The analysis and conclusions set forthare those of the authors and do not indicate concurrence by other members of the research staff or theBoard of Governors. References in publications to the Finance and Economics Discussion Series (other thanacknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.
Inter-firm Relationships and Asset Prices
Carlos Ramırez∗
January 2017
ABSTRACT
This paper proposes a novel link between the propagation of shocks within production networksand asset prices. It develops a dynamic network model in which the propagation of firm cash-flowshocks via inter-firm relationships affects the economy’s equilibrium asset prices. When calibratedto match key features of customer–supplier networks in the United States, the model generateslong-run risks, high and volatile risk premia, and a low and stable risk-free rate. Consistent withdata from firms in manufacturing and service industries, the model predicts that central firms in thenetwork command lower risk premiums than peripheral firms, and that firm-level return volatilitiesexhibit a high degree of co-movement.
Keywords: Inter-firm Relationships, Shock Propagation, Networks, Equilibrium Asset Prices.
JEL classification: G12, E32, L10.
∗Board of Governors of the Federal Reserve System. I thank Fernando Anjos; Celso Brunetti; Elena Carletti;Francisco Cisternas; Nathan Foley-Fisher; Nicola Gennaioli; Stefan Gissler; Brent Glover; Richard Green; AnishaGhosh; Benjamin Holcblat; Steve Karolyi; Robert Kieschnick; Yongjin Kim; Mete Kilic; Ian Kotliar; Jun Li; BorghanNarajabad; Artem Neklyudov; Kim Peijnenburg; Silvio Petriconi; Valery Polkovnichenko; Fulvio Ortu; Emilio Osam-bela; Doriana Ruffino; Stefano Sacchetto; Alessio Saretto; Julien Sauvagnat; Fabiano Schivardi; Duane Seppi; ChesterSpatt; Claudio Tebaldi; Stephane Verani; Hannes Wagner; Malcolm Wardlaw; Ariel Zetlin-Jones; and seminar partici-pants at the 15th Trans-Atlantic Doctoral Conference at LBS, Carnegie Mellon, INFORMS, Bocconi, IESE, Universityof Texas at Dallas, Federal Reserve Board, Cornerstone Research, Central Bank of Chile, Portsmouth-Fordham Con-ference in Banking and Finance, the 5th CIRANO-Walton Conference on Networks, Luxembourg School of Finance,and ASSET 2016 for their valuable suggestions. I am especially grateful to Burton Hollifield, Bryan Routledge andR. Ravi for their helpful discussions. All remaining errors are my own. This article represents the view of the author,and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System orother members of its staff. E-mail: [email protected].
Inter-firm relationships, such as strategic alliances, joint ventures, research and development (R&D)
partnerships, and customer–supplier relationships, are ubiquitous in modern economies. A growing
body of empirical work emphasizes the importance of inter-firm relationships in the case of firms’
distress and shows that they may serve as propagation mechanisms of negative shocks to individual
firms.1 As firm-level shocks may spread to a non-negligible fraction of firms in the economy, such
a shock propagation can, in principle, generate aggregate fluctuations, altering equilibrium asset
prices and aggregate risk premia.2 Despite the importance of this shock propagation, its equilibrium
asset pricing implications have been overlooked in the literature.3 This paper fills this gap by
developing a model that links equilibrium asset prices and aggregate risk premia to the propagation
of cash-flow shocks across firms in a network economy.
When calibrated to match key characteristics of customer–supplier networks in the United
States, I show that a dynamic network economy model in which cash-flow shocks propagate via
inter-firm relationships and investors have Epstein-Zin-Weil preferences generates long-run risks
and quantitatively replicates prime characteristics of asset market data. In particular, the model
generates high and volatile risk premia, as well as a low and stable risk-free rate. Moreover, the
model predicts that central firms in the network command lower risk premiums than peripheral
firms, and that firm-level return volatilities exhibit a high degree of co-movement. These predictions
are consistent with data from firms in manufacturing and service industries but cannot be accounted
for by standard asset pricing models.
The model has two main features. First, firms’ cash-flow growth prospects are determined via
an exogenous network of long-term inter-firm relationships, as firm-level cash-flow shocks propagate
via such relationships.4 In the model, the propensity of inter-firm relationships to transmit shocks
varies over time to capture changes in relationship-specific characteristics that make connected
firms more vulnerable to negative spillovers. Changes in relationship-specific characteristics may
1See Hertzel et al. (2008), Jorion and Zhang (2009), Boone and Ivanov (2012), Carvalho, Nirei, and Saito (2014),Boyarchenko and Costello (2015), Todo, Nakajima, and Matous (2015), Boehm, Flaaen, and Pandalai-Nayar (2015)and Barrot and Sauvagnat (2016) among others.
2Using French firm-level data from 1990 to 2007, Di Giovanni, Levchenko, and Mejean (2014) provide empiricalevidence of the importance of firm-specific shocks in generating aggregate fluctuations.
3The contemporaneous work of Herskovic (2015) is an exception.4Long-term inter-firm relationships may allow connected firms to circumvent difficulties in contracting due to
unforeseen contingencies, asymmetries of information, and specificity on firms’ investments, e.g., Williamson (1979,1983). The network is assumed to be exogenous because the focus of this paper is on the effect of shock propagationon equilibrium asset prices rather than on strategic network formation. For endogenous formation of productionnetworks, see Oberfield (2013), Chaney (2014, 2016), and Lim (2016).
2
arise from variation in restrictions on firms’ use of alternative inputs, e.g. Barrot and Sauvagnat
(2016). Variation in such restrictions may emerge from changes—driven by innovation or indus-
try regulation—in production technologies, complementarities among firms’ activities, or market
competition. As the propensity of inter-firm relationships to transmit shocks varies over time, the
mechanism by which shocks propagate across firms also varies over time, introducing an endogenous
time-varying correlation structure among firms’ cash-flows. Second, investors have Epstein-Zin-Weil
preferences—as in standard asset pricing models—and, thus, care about uncertainty regarding firms’
long-term growth prospects.
In the calibrated model, low-frequency changes in the shock propagation mechanism endoge-
nously generate persistence in firms’ growth prospects due to the long-term nature of inter-firm
relationships. In equilibrium, the persistence in firms’ growth prospects drives a small and per-
sistent component in expected aggregate consumption growth. Because investors have recursive
preferences, the model accounts for sizeable risk premiums and a small and stable risk-free rate, as
low-frequency movements in consumption growth induce large movements in marginal utility and
stock prices. Intuitively, sizeable risk premiums arise because investors fear that extended periods
of low economic growth coincide with low asset prices. Likewise, a small risk-free rate is driven by
investors saving for long periods of low economic growth.
In the cross section, the calibrated model predicts that central firms in the network command
lower risk premiums than peripheral firms. Central firms tend to benefit from the diversification
of their customers and suppliers and, thus, mitigate contagion risk better than less central firms,
commanding lower risk premiums. Consistent with data from firms in manufacturing and service
industries, the model generates a realistic annual return spread of 3.8% between firms in the lowest
tercile of centrality and firms in the highest tercile of centrality.5 This return spread, which cannot
be accounted for by standard asset pricing models, arises naturally in equilibrium as compensation
for contagion risk.
In the time series, the calibrated model predicts high co-movement in firm-level return volatilities
as changes in the shock propagation mechanism drive fluctuations in growth opportunities and
uncertainty across firms. These cross-sectional fluctuations translate into simultaneous changes in
5Wu and Birge (2014) provide complementary evidence that manufacturing firms that are more central in thenetwork earn lower returns.
3
stock returns, which generate a factor structure in the time series of firm-level returns and returns’
volatilities.6
The small and persistent component in expected consumption growth generated by low-frequency
movements in the shock propagation mechanism provides an equilibrium foundation for long-run
risk models in the spirit of Bansal and Yaron (2004). Moreover, the model helps explain the cross
section of expected returns as it provides a mapping between firms’ importance in the network and
their contagion risk. The model suggests that extending standard asset pricing models to take into
account the way that shocks propagate within production networks can make significant progress
toward generating a unifying framework that simultaneously captures dynamics of the aggregate
and the cross section of stock returns.
This paper contributes to three strands of the literature. First, the paper develops a new
theoretical framework that adds to a growing body of work focused on understanding the effects
of economic linkages in asset pricing properties, e.g., Buraschi and Porchia (2012), Ahern (2013),
and Herskovic (2015).7 Unlike these papers, my model emphasizes relationships at the firm level
to explore the asset pricing properties that stem from the propagation of shocks within production
networks.
Second, this paper adds to a body of work that explores how granular shocks may lead to
aggregate fluctuations in the presence of linkages among different sectors of the economy, e.g.,
Carvalho (2010), Gabaix (2011), Acemoglu et al. (2012); Acemoglu, Ozdaglar, and Tahbaz-Salehi
(2015), Oberfield (2013), Carvalho and Gabaix (2013), Blume et al. (2013), Elliott, Golub, and
Jackson (2014), Chaney (2014, 2016), and Lim (2016). This paper contributes to this literature by
exploring the asset pricing implications of linkages at the firm level and studying how changes in
the propagation of shocks within a network economy affect not only aggregate variables but also
equilibrium asset prices and aggregate risk premia.
Third, this paper adds to recent research that examines the potential sources of long-run risks,
6The factor structure in returns’ volatilities is aligned with recent empirical evidence documented by Duarte et al.(2014).
7Buraschi and Porchia (2012) show that more central firms in a market-based network have lower price dividendratios and higher expected returns. Using the network of intersectoral trade, Ahern (2013) provides evidence thatfirms in more central industries have greater exposure to systematic risk. Herskovic (2015) focuses on efficiency gainsthat come from changes in the input-output network and how those changes are priced in equilibrium. My paper, onthe other hand, focuses on how changes in the propagation of shocks within a fixed network alter equilibrium assetprices and risk premia.
4
e.g., Kaltenbrunner and Lochstoer (2010), Kung and Schmid (2015), Bidder and Dew-Becker
(2016).8 This paper contributes to this literature by showing that long-run risks may also arise
from changes in the way that shocks propagate within sticky production networks.
The rest of the paper is organized as follows. Section I introduces the baseline model. Section
II describes aggregate consumption growth in the baseline model. Section III derives expressions
for the market return, the risk-free rate, the price of risk, and firms’ equilibrium asset prices and
expected returns. Section IV uses data on customer–supplier networks in the United States to
calibrate the model. Section V shows that changes in the propagation mechanism of shocks in
customer–supplier networks are quantitatively important to understand variations in stock returns
in both the aggregate and the cross section. Section VI concludes.
I. Baseline Model
The baseline model embeds a single-good dynamic endowment economy, in the spirit of Lucas
(1978), into a standard asset pricing model with investors with Epstein-Zin-Weil preferences. In an
otherwise standard dynamic endowment economy, the outputs of the economy’s productive units,
henceforth firms, are determined by a network of long-term inter-firm relationships. The single-
good endowment economy framework is assumed to facilitate exposition and can be extended to a
production network setting in which every firm produces a different good and each good is necessary
to produce other goods in the economy.9
A. The environment
Consider an economy with one perishable good and an infinite time horizon. Time is discrete
and indexed by t ∈ 0, 1, 2, · · · . In each period, the single good is produced by n infinitely lived
firms, with n being potentially large. The economy is populated by a large number of identical,
infinitely lived individuals who are aggregated into a representative investor with Epstein-Zin-Weil
8Kaltenbrunner and Lochstoer (2010) shows that long-run risks endogenously arise in a standard productioneconomy model, even when technology growth is i.i.d., because of consumption smoothing. Kung and Schmid (2015)shows that a model of endogenous innovation and R&D is able to generates long-run risks, while Bidder and Dew-Becker (2016) shows that long-run risks arise in an economy in which investors are pessimistic and not sure aboutthe true model driving the economy.
9In such an environment, equilibrium asset prices are similar to the ones obtained here if the representativeinvestor has preferences over a particular basket of goods, e.g., Herskovic (2015).
5
preferences who owns all assets in the economy. Firms’ outputs, henceforth cash flows, are related
via a network of inter-firm relationships. The production network is described by a graph consisting
of a set of nodes, which represent firms, together with edges joining certain pairs of nodes, which
represent inter-firm relationships. To fix notation, let Gn denote the production network among n
firms. Because I focus on the effect of Gn on asset prices rather than on the strategic formation
of inter-firm relationships, these relationships are assumed to be exogenously determined and fixed
before t = 0.10
B. The network of inter-firm relationships and firms’ cash flows
Firms’ cash flows vary stochastically over time and depend not only on the production network
Gn but also on the way that cash-flow shocks propagate within Gn. In particular, relationships
generate benefits—such as information and resource sharing, access to new markets, and easing
of financial constraints via trade credit—which increase a firm’s cash-flow growth rate. However,
relationships also increase a firm’s exposure to negative cash-flow shocks that affect other firms, as
negative cash-flow shocks spread through probabilistic contagion via relationships. Such a trade
off is captured by the following reduced-form equation:
log
(yi,tYt−1
)≡ α0 + α1di − α2
√nεi,t , i ∈ 1, · · · , n , (1)
where yi,t denotes firm i’s cash flow at period t, and Yt−1 denotes the aggregate output of the
economy at t− 1. Parameters α0, α1, and α2 are non-negative and equal across firms. Parameter
di represents the number of direct relationships of firm i—which may differ across firms. The
term√n is included in equation (1) as a normalization factor to help characterize the equilibrium
distribution of aggregate consumption growth later on. Uncertainty in yi,t is introduced by a
Bernoulli random variable εi,t, which equals one if firm i faces a negative cash-flow shock at period
t and zero otherwise. Because
log
(yi,tYt−1
)= log
(yi,tyi,t−1
)+ log
(yi,t−1
Yt−1
),
10See Demange and Wooders (2005), Goyal (2007), and Jackson (2008) for a detailed description of networkformation models. For models of endogenous formation of production networks, see Oberfield (2013), Chaney (2014,2016), and Lim (2016), among others.
6
parameter α2 in equation (1) measures the decrease in a firm’s cash-flow growth if a firm faces a
negative cash-flow shock. Thus, transitory negative cash-flow shocks have an enduring effect as they
depress not only firms’ cash-flows today but also firms’ growth prospects. Parameter α1 captures
the increase in a firm’s cash-flow growth due to each direct relationship, whereas α0 captures the
part of firms’ cash-flow growth that is unrelated to benefits or costs associated with inter-firm
relationships.11
The distribution of εi,t is determined by the following stochastic process—which simplifies the
modeling and abstracts from the temporal propagation of shocks. At the beginning of period
t, each firm may face a negative cash-flow shock, independently of other firms, with probability
0 < q < 1—which is equal across firms and time invariant. Firms affected by idiosyncratic cash-
flow shocks have a probability of causing their direct partners to face a negative cash-flow shock as
well, allowing negative shocks to potentially continue to spread from the newly affected firms.12 In
particular, a negative cash-flow shock to firm i at period t also affects firm j at period t, and, thus,
εi,t = εj,t = 1 if two things happen: (1) there exists a sequence of relationships that connects firms
i and j in Gn and (2) each relationship in that sequence is active in transmitting shocks at period
t. For simplicity, in each period each relationship is either active in transmitting cash-flow shocks
or not—independently of all other relationships. The relationship between firms i and j is active
in transmitting shocks at period t with probability pijt. The production network Gn is assumed to
be undirected, and, thus, pijt = pjit, ∀(i, j) ∈ Gn, ∀ t.13
11In the absence of relationships, α0 equals the growth rate of the economy if Yt ≡∏n
i=1 y1/ni,t .
12Within the baseline model, only negative shocks are allowed to propagate in a probabilistic manner to focus onthe effect on asset prices of the propagation of shocks in case of firms’ distress. However, the baseline model can beeasily extended to allow positive and negative cash-flow shocks to propagate through the network. To do so, defineψi,t ≡ εi,t − 1/2 so that cash-flow shocks can be positive and negative. Then, redefine equation (1) as
log
(yi,tYt−1
)= α0 + α1di − α2
√nψi,t
= α0 + α2
√n/2︸ ︷︷ ︸+α1di − α2
√nεi,t
= α0 + α1di − α2
√nεi,t ,
which is analogous to equation (1). The cross-sectional results in this paper continue to hold as long as the decreasein firms’ cash-flow growth due to negative shocks is larger than the increase in firms’ cash-flow growth due to positiveshocks.
13This stochastic process can be thought of as a variation of either a reliability network or a bond percolationmodel in each period. In a typical reliability network model, the edges of a given network are independently removedwith some probability. The remaining edges are assumed to transmit a message. A message from node i to j istransmitted as long as there is at least one path from i to j after edge removal—see Colbourn (1987) for more details.Similarly, in a bond percolation model, edges of a given network are removed at random with some probability. Edgesthat are not removed are assumed to percolate a liquid. The question in percolation is whether the liquid percolates
7
The value of pijt measures the propensity of the relation (i, j) to transmit cash-flow shocks from
firm i (j) to j (i) at period t. At a fundamental level, the value of pijt captures relationship-specific
characteristics that make firm i (j) more vulnerable to negative spillovers coming from firm j (i),
which cannot be mitigated through contractual protections at period t. Intuitively, the higher the
value of pijt, the higher the likelihood that problems affecting firm i (j) also affect firm j (i) at
period t. In the context of supply chains, pijt may capture restrictions on firm i’s and j’s use of
alternative inputs at period t. The higher the value of pijt, the higher the switching costs firms i or
j may face at period t, and, thus, the higher the likelihood that a negative cash-flow shock to firm
i (j) also affects firm j (i), provided that firm j (i) may not be able to restructure its production
sufficiently fast to overcome firm i (j)’s negative cash-flow shock.
Probabilities pijt(i,j)∈Gnare drawn from a Beta distribution with parameters ζ1t > 0 and
ζ2t > 0 at the beginning of period t. Parameters ζit > 0, i = 1, 2, which are drawn at the very
beginning of period t, determine the shape of the distribution of propensities across relationships
at period t. The model timeline at period t is depicted in figure 1.
ζ1t and ζ2tare drawn
pijt(i,j) ∈ Gn
are drawn from β (ζ1t, ζ2t)
Relationships thattransmit shocksare determined
Firm-specific shocksare realized
Shockspropagate
Period t
Figure 1. Model timeline in period t.
To sum up, equation (1) captures some of the potential consequences of long-term inter-firm
relationships in a simple manner. While inter-firm relationships may increase firms’ growth op-
portunities via efficiency gains, these relationships may also have additional consequences as they
from one node to another in the network—which is similar to the problem of transmitting a message in a reliabilitycontext. For more details see Grimmett (1989), Stauffer and Aharony (1994), and Newman (2010, Chapter 16.1).Blume et al. (2013) analyze a propagation mechanism similar to the one analyzed here. They focus, however, onstrategic network formation issues in a static environment. They provide asymptotic bounds on the welfare of bothoptimal and stable networks and show that small amounts of “over-linking” may impose large losses in welfare tonetworks’ participants.
8
increase a firm’s exposure to negative cash-flow shocks that affect a broader set of firms in the
economy. Although equation (1) is a reduced-form formulation, it can be recast so it is generated
within an equilibrium context. For instance, Goyal and Moraga-Gonzalez (2001) obtain similar
dynamics for firms’ profits in strategic environments where firms collaborate in an R&D network
to decrease their production costs, but they also compete with their collaborators within the same
homogeneous good market.
C. Shock propagation and the cross-sectional distribution of εi,t
Given how shocks propagate, the joint distribution of the cross-sectional sequence εi,tni=1 is
determined by Gn, q, and the process driving the stochastic propensity matrix pt ≡ [pijt](i,j) ∈ Gn.
Moreover, the marginal distribution of εi,t, conditional on pt, depends on q, the network Gn, and
the location of firm i in Gn. In other words,
P(εi,t = 1
∣∣pt)
= f (q,Gn, location of firm i in Gn) , (2)
where P(εi,t = 0
∣∣pt)= 1 − P
(εi,t = 1
∣∣pt), and f(·) is a mapping characterized by the stochastic
process described in section I.B—which endogenously generates a time-varying correlation structure
among firms’ cash flows as pt varies over time.
Despite the fact that the mapping f(·) is hard to characterize for large n, its properties are
easy to describe given the formulation of the stochastic process that generates it. First, in the
absence of relationships, P(εi,t = 1
∣∣pt)= P (εi,t = 1) = q , ∀ i and ∀ t, so cash-flow growth rates
are independent and identically distributed across firms over time. Second, if only one sequence of
relationships exists between two firms, the longer the sequence, the smaller the correlation between
their cash-flow growth rates. Thus, in network economies in which there is at most one sequence
of relationships between any two firms, the more distant the two firms are, the less related their
cash flows.14
14Having this feature—which is sometimes called correlation decay, e.g., Gamarnik (2013)—helps a great deal toobtain numerical solutions of the model relatively fast when n is large.
9
D. Temporal changes in shock propagation within the network economy
To capture temporal changes in relationship-specific characteristics, the shape parameters ζit,
i = 1, 2, are allowed to vary over time. Variation in the shape parameters may arise from
changes in firms’ production technologies, complementarities among firms’ activities, or market
competition. For simplicity, ζit can take two values, ζiL or ζiH , with ζiL < ζiH , i = 1, 2, and the
shape parameter vector ζt ≡ [ζ1t ζ2t] follows a four-state ergodic Markov process with transition
matrix Ω and states ζLL ≡ [ζ1L ζ2L], ζLH ≡ [ζ1L ζ2H ], ζHL ≡ [ζ1H ζ2L], and ζHH ≡ [ζ1H ζ2H ].
II. Distribution of Consumption Growth
Two features of the model are important to understand aggregate consumption growth: (a)
the topology of the production network Gn and (b) how shocks propagate across firms, captured
by the propensity matrix pt and its dynamics. In this section, I study how changes in these two
features affect the distribution of aggregate consumption growth and, thus, alter the distribution
of the pricing kernel. Let ∆ct+1 ≡ log(Ct+1
Ct
)and xt+1 ≡ log
(Yt+1
Yt
)denote the log consumption
and output growth at t+ 1, respectively. In equilibrium, ∆ct+1 = xt+1. For tractability, consider
Yt ≡∏n
i=1 y1/ni,t . Then, it follows from equation (1) that
∆ct+1 = xt+1 = log
(n∏
i=1
(yi,t+1
Yt
)1/n)
=
n∑
i=1
1
nlog
(yi,t+1
Yt
)
= α0 + α1
(1
n
n∑
i=1
di
)
︸ ︷︷ ︸−α2
√n
(1
n
n∑
i=1
εi,t+1
)
︸ ︷︷ ︸
= α0 + α1 d − α2
√n Wn,t+1 , (3)
where d denotes the average number of relationships per firm in the economy, whereas Wn,t+1
denotes the average number of firms affected by negative cash-flow shocks at period t+1. It follows
from equation (3) that the distribution of ∆ct+1 is determined by the distribution of√nWn,t+1.
Because the distribution of√nWn,t+1 is affected by pt+1 and the topology of Gn, these two features
10
also affect the distribution of ∆ct+1.15
To appreciate the importance of pt+1 and the topology of Gn in determining the distribution
of ∆ct+1, consider two cases. First, suppose there are no relationships. In this case, εi,t+1ni=1 is
a sequence of i.i.d. Bernoulli random variables and nWn,t+1 follows a Binomial distribution. By
the Central Limit Theorem (CLT),√nWn,t+1 is normally distributed as n grows large. Provided
the absence of relationships, the realization of the matrix pt+1 is irrelevant to determining the
distribution of ∆ct+1, as the unconditional mean and variance of Wn,t+1 are q and q(1−q)n , respec-
tively. Second, suppose that all firms have two relationships and that the propensity to transmit
shocks of each relationship is p, which does not vary over time. Then, εi,t+1ni=1 is a sequence of
dependent Bernoulli random variables and nWn,t+1 follows approximately a Binomial distribution
if p is sufficiently small. In this case, the propensity of relationships to transmit shocks, p, affects
the distribution of consumption growth, as the unconditional mean and variance of Wn,t+1 are
approximately π and π(1−π)n , respectively; where π ∈ [0, 1] solves the following equation:
π = q + (1− q)πp (πp+ 2 [p(1− π) + π(1− p)]) .
Thus, in the presence of relationships, pt+1 and the topology of Gn affect the distribution of
consumption growth, as εi,t+1ni=1 is a sequence of dependent Bernoulli random variables. In this
case, the conditions under which a CLT holds may not be satisfied as n grows large.16 Relationships
may generate convoluted interdependencies among firms’ cash flows, which makes it difficult to
characterize the distribution of ∆ct+1. Figure 2 illustrates the previous point. Figure 2(a) depicts
a star network in an economy with n = 5 firms, whereas figure 2(b) depicts the empirical probability
density function of√nWn,t+1 for the star network depicted in figure 2(a). As figure 2(b) shows,
15The definition of Yt implies that positive aggregate production requires positive production by each firm. Toassume that Yt ≡
∏ni=1 y
1/ni,t is similar to assuming that Yt is proportional to
∑ni=1 yi,t if n is sufficiently large and all
yi,t 6= 0. The argument follows from applying a first order Taylor series expansion to log (Yt) in which aggregate output,
Yt ≡∑n
i=1 yi,t. A different way of justifying that Yt ≡∏n
i=1 y1/ni,t is to consider that every firm produces a different
perishable good and each good is necessary to produce other goods in the economy. In such an environment, oneobtains asset pricing properties similar to the ones obtained in this paper if the representative investor has preferencesover a Cobb-Douglas consumption aggregator of the form Ct ≡
∏ni=1 c
1/ni,t , where ci,t represents consumption of the
good produced by firm i at time t.16For a large variety of network topologies, simulation shows that the distribution of ∆ct+1 may differ from
a normal distribution. In particular, if some elements of the matrix pt+1 are sufficiently close to one and Gn islocally connected—i.e., there is at least one sequence of relationships between any two firms in an arbitrarily largeneighborhood around any given firm—then a non-negligible fraction of firms in the economy are almost surely affectedby negative cash-flow shocks. Therefore, the distribution of ∆ct+1 may exhibit thicker tails than a normal distributionwould.
11
the distribution of√nWn,t+1 may differ from a normal distribution if the elements of the matrix
pt+1 are sufficiently close to 1. In particular, as some components in pt+1 tend toward one, the
distribution of√nWn,t+1 tends to be bimodal.
Despite the existence of relationships and the convoluted dependencies they may generate among
firms’ cash flows, the topology of Gn and values in the matrix pt+1 can be restricted so that ∆ct+1 is
normally distributed as n grows large (see Appendix A). In such a case, keeping track of temporal
changes of the distribution of ∆ct+1 is equivalent to keeping track of temporal changes in averages
and standard deviations. In particular, if shocks tend to remain locally confined—i.e., no shock
propagates over a large fraction of firms in the economy—εi,t+1ni=1 becomes a sequence of weakly
dependent random variables to which a CLT can be applied. Then, the dynamics of consumption
growth can be recast as a version of Hamilton (1989)’s Markov-switching model.
III. Equilibrium Asset Prices
To see what the production network Gn and the dynamic of ζt imply for equilibrium asset prices,
I embed the cash-flows correlation structure that is endogenously generated by the baseline model
in a standard asset pricing framework. The representative investor has Epstein-Zin-Weil recursive
preferences to account for asset pricing phenomena that are challenging to address with power
utility preferences. The asset pricing restrictions on the gross return of firm i, Ri,t+1, are
Et
(Mt+1Ri,t+1
)= 1 , (4)
where Mt+1 ≡[β(e∆ct+1
)−ρ] 1−γ
1−ρ[Ra,t+1
] 1−γ1−ρ
−1represents the pricing kernel at t + 1 and Ra,t+1
denotes the gross return on aggregate wealth—an asset that delivers aggregate consumption as its
dividend each period. Parameter ρ > 0, ρ 6= 1, represents the inverse of the inter-temporal elasticity
of substitution (IES), γ > 0 is the coefficient of relative risk aversion for static gambles, and β > 0
measures the subjective discount factor under certainty.17
To solve the model, I look for equilibrium asset prices so that price–dividend ratios are stationary,
17If γ = ρ, these recursive preferences collapse to the standard case of Von Neumann-Morgenstern (VNM) time-additive expected utility. The functional form of the pricing kernel when ρ = 1 is different from the one shown above.See Weil (1989, Appendix A) for details. I use the standard terminology to describe γ and ρ. However, Garcia,Renault, and Semenov (2006) and Hansen et al. (2007) indicate that this interpretation may not be correct if ρ 6= γ.
12
as in Mehra and Prescott (1985), Weil (1989), and Kandel and Stambaugh (1991), among many
others. Because equilibrium values are time-invariant functions of the state of the economy—which
is determined by the state of the parameter vector ζt—the index t can be eliminated. Hereinafter, c
denotes the current level of aggregate consumption, y denotes the current level of aggregate output,
and s ∈ LL,LH,HL,HH denotes the current state of parameter vector ζ.
I first solve for the price of aggregate wealth and the risk-free rate. These expressions are then
used to solve for equilibrium asset prices and expected excess returns in the cross section.
PROPOSITION 1 (Price of Aggregate Wealth): Let Pa(c, s) denote the current price of aggregate
wealth. Pa(c, s) = wasc, where wa
s is the solution of the following nonlinear system of equations,
was = β
∑
s′∈LL,LH,HL,HH
ωs,s′E
(e(1−γ)∆ct+1
∣∣s′)(wa
s′ + 1)1−γ1−ρ
1−ρ1−γ
, (5)
where E(·∣∣s′)denotes the conditional expectation operator if the state of ζ is s′ and ωs,s′ represents
the (s, s′) element of Ω.
I restrict my analysis to the set of model primitives in which the existence of a non-negative
solution of the system of equations (5) is ensured.18 The expected period gross return of aggregate
wealth in the current state is then
E (Ra|s) =∑
s′∈LL,LH,HL,HH
ωs,s′was′ + 1
was
E
(e∆ct+1
∣∣s′). (6)
It follows from equations (5) and (6) that the price and expected period return of aggregate wealth
are driven by: (a) the topology of the production network Gn, and (b) the dynamics of ζ. As
noticed in section II, the topology of Gn and pt+1 drive the distribution of ∆ct+1. Given a network
Gn, ζ shapes the distribution of aggregate consumption growth, as ζ determines pt+1. In particular,
18Provided that e∆ct is positive for all t, parameters ρ and γ need to be restricted so that the function h(·) definedas
h (wai ) ≡ β
∑
j∈LL,LH,HL,HHωi,jE
(e(1−γ)∆ct+1
∣∣j)(wa
j + 1)1−γ1−ρ
1−ρ1−γ
is continuous. If h(·) is continuous, the system of equations (5) has a solution by Brouwer’s Fixed Point Theorem.Further restrictions in the set of parameter values can be imposed such that the solution of the system of equationsis unique.
13
changes in ζ convey changes in the cross-sectional distribution of pij(i,j) ∈ Gnwhich determines
how shock propagate across firms. The dynamics of ζ—parameterized by Ω—affect the price and
the expected period return of aggregate wealth, as Ω determines (a) how frequently the economy is
in a state in which relationships are more prone, on average, to transmit negative cash-flow shocks
and (b) how persistent are changes in the shock propagation mechanism.
I next consider the risk-free asset, which pays one unit of the consumption good during the next
period with certainty.
PROPOSITION 2 (Risk-free Rate): Let Rf (s) denote the period gross return of the risk-free asset
in the current state. Rf (s) solves
1
Rf (s)= β
1−γ1−ρ
∑
s′∈LL,LH,HL,HH
ωs,s′E
(e−γ∆ct+1
∣∣s′)(wa
s′ + 1
was
) ρ−γ1−ρ
, (7)
where was are the solutions of the system of equations (5).
It follows from equation (7) that the equilibrium risk-free rate is also driven by the topology of
Gn and the dynamics of ζ, as these two features affect the distribution of aggregate consumption
growth and prices of aggregate wealth.
Using the previous expressions, I now study what the production network Gn and dynamics of ζ
imply for the cross section of asset prices and risk premiums. The following proposition determines
the ex-dividend stock price of firm i and its expected period return.
PROPOSITION 3 (Firms’ Stock Prices and Expected Period Returns): Let Pi(y, s) denote the
current ex-dividend stock price of an asset that delivers firm i’s cash flows as its dividend each
period. For large n, Pi(y, s) ≈ vi(s)y, where vi(s) is the solution of the following linear system of
equations
vi(s) = β1−γ1−ρ
∑
s′∈LL,LH,HL,HH
ωs,s′
(was′ + 1
was
) ρ−γ1−ρ
E
(ext+1−γ∆ct+1
∣∣s′)vi(s
′)
(8)
+ β1−γ1−ρ eα0+α1di
∑
s′∈LL,LH,HL,HH
ωs,s′
(was′ + 1
was
) ρ−γ1−ρ
E
(e−γ∆ct+1
∣∣s′) [
1− πi(s′)] ,
where πi(s′) ≡ E
(εi,t+1
∣∣s′)= P
(εi,t+1 = 1
∣∣s′). Moreover, the expected one period gross return of
14
firm i is given by
E
(Ri,t+1
∣∣s)
=1
vi(s)
∑
s′∈LL,LH,HL,HH
ωs,s′
vi(s
′)E(ext+1
∣∣s′)+ eα0+α1di
(1− πi(s
′)) . (9)
To appreciate the importance of the location of firm i in Gn on equilibrium asset prices and
expected returns, suppose all firms have the same number of direct relationships. Then, di = d
and πi = π ≥ q for all i. It follows from the second term in the right hand side of (8) that all firms
have the same ex-dividend stock price. As equation (8) shows, differences in prices across firms
arise solely from differences in the location of firms in Gn. In particular, differences in prices across
firms are driven not only by the number of direct relationships of a firm, captured by di, but also
by the set of firms to which a firm is connected, captured by πi. The same applies for the cross
section of expected returns. Differences in expected returns across firms arise solely from differences
across firms’ locations. To understand the cross section of firms’ risk premiums, equation (4) can
be rewritten as a beta pricing model,
E
(Ri,t+1
∣∣s)−Rf (s) =
Cov
(Ri,t+1, Mt+1
∣∣s)
Var(Mt+1
∣∣s)
︸ ︷︷ ︸
−Var
(Mt+1
∣∣s)
E
(Mt+1
∣∣s)
︸ ︷︷ ︸
, (10)
βi,M
(s) λM(s)
where βi,M
(s) and λM(s) denote firm i’s quantity of risk and the conditional price of risk in state
s, respectively. The following proposition determines the conditional price of risk, λM(s).
PROPOSITION 4 (Conditional Price of Risk: λM(s)): The conditional price of risk in state s,
λM(s), equals
λM(s) =
1
Rf (s)−Rf (s)
β
2(
1−γ1−ρ
) ∑
s′∈LL,LH,HL,HH
ωs,s′
(was′ + 1
was
)2(
ρ−γ1−ρ
)
E
(e−2γ∆ct+1
∣∣s′) , (11)
where Rf (s) denotes the period gross return of the risk-free asset in state s.
As equation (11) shows, the price of risk is time varying. Changes in ζ introduce changes in
the cross-sectional distribution of pij(i,j)∈Gn, which, in turn, generate changes in the distribution
15
of aggregate consumption growth, in the price of aggregate wealth, and in the risk-free rate, all of
which manifest in changes of the price of risk. To compute firms’ quantities of risk, equation (10)
can be rearranged as
βi,M
(s) =E
(Ri,t+1
∣∣s)−Rf (s)
λM(s)
(12)
so that βi,M
(s) can be computed from equations (7), (9), and (11).
IV. Calibration
So far, the model illustrates how changes in the propagation mechanism of shocks within a
production network alters equilibrium asset prices and expected returns. I now calibrate the model
to match several features of customer–supplier networks in the United States and explore its ability
to replicate characteristics of asset returns. Section IV.A describes the data and the strategy
employed to calibrate the network Gn as well as the dynamics of ζ. Section IV.B describes the
selection of the rest of the parameters in the model.
A. Description of Data, Customer–Supplier Networks, and Dynamics of ζt
I use annual data on customer–supplier relationships among public U.S. firms to pin down the
topology of Gn. The Statement of Financial Accounting Standards (SFAS) No.131 requires firms
to report the existence of customers who represent more than 10% of their annual sales. This
information is available on the COMPUSTAT Segment files. However, these files tend to list only
abbreviations of customers’ names. I then use the Cohen and Frazzini (2008) database on customer–
supplier relationships—a subset of the COMPUSTAT Segment database—in which firms’ principal
customers are uniquely identified. Their dataset consists of 6,425 different public firms, considers
common stocks, and represents 26,781 unique annual customer–supplier relationships from 1980
to 2005. Customer–supplier relationships last about three years on average. The distribution
of firms’ sizes resembles the size distribution of the CRSP universe over the sample period, but
the size distribution of firms’ principal customers is tilted toward large companies, as firms are
only required to report customers that represent more than 10% of their annual sales. Table II
16
reports the distribution of firms across major industry groups for which monthly return data are
available from 1980 to 2004. As table II shows, almost 70% of companies in the dataset are either
manufacturing or service firms.19
Using the Cohen and Frazzini (2008) database, I construct networks at an annual frequency over
the sample period. For the network of year t, nodes represent public firms that report customer–
supplier agreements at year t and links represent customer–supplier relationships during that year.
The idea is to select a topology for the benchmark economy that matches several features of the
time series of U.S. customer–supplier networks. For illustration, figures 4, 5, and 6 depict the
U.S. customer–supplier networks from 1980 to 1997. As these figures show, U.S. customer–supplier
networks are highly asymmetric in the sense that only a few firms are connected to many others,
while most firms have either one or at most two connections. In fact, the degree distributions of
these networks, which measure the frequency of firms with a given number of direct relationships,
are highly right-skewed and their upper tails can be approximated via power law distributions.
Table III shows averages and standard deviations for key characteristics of the time series of U.S.
customer–supplier networks. As Table III shows, the high asymmetry of U.S. customer–supplier
networks is fairly persistent over the sample period, as measured by the ratio
SD(exponent of power law distribution fitted to degree distribution)
Mean(exponent of power law distribution fitted to degree distribution)= 0.07.
I select the topology of the benchmark economy by manually constructing a network whose
topology simultaneously matches several of the averages reported in Table III. In particular, the
topology of the benchmark economy matches the average number of firms, and the average empirical
degree distribution of the time series of U.S. customer–supplier networks. The selected topology
also captures the clustering pattern exhibited by these networks. Using two consecutive depth-first
searches, I compute the size of the five largest connected components in each customer–supplier
network. A connected component is a subset of the network in which any two firms are connected
to each other by sequences of relationships, and which is connected to no additional firms in the
network. The selected topology matches the average size of each of the five largest connected
19This data is available at http://www.econ.yale.edu/∼ af227/. According to the U.S. Department of Labor,manufacturing firms (division D) include companies engaged in mechanical or chemical transformation of material orsubstances into new products, whereas service firms (division I) are companies that usually provide a wide variety ofservices for individuals, businesses, and governments, such as hotels, automobile repair, and health services.
17
components over the sample period.20 I also restrict the topology of Gn to have no cycles so that
firms’ probabilities of facing negative shocks in each state of the economy—expressions πi(s)ni=1
in equations (8) and (9)—are easy to compute.21 This restriction seems to be innocuous, because
cycles are not frequent in the dataset. Figure 7 depicts the degree distribution of the benchmark
production network.
To pin down the parameters that define the dynamics of ζ, I need proxies for the cross-sectional
distributions of propensities of relationships to transmit shocks. However, the propensity of the
relationship between firm i and j at year t, pijt, is unobservable. In the context of supply chains,
nonetheless, evidence documented by Barrot and Sauvagnat (2016) suggest that pijt is related with
firms i and j’s input specificities. As firms’ input specificities are likely to depend on the percentage
of sales that a customer represents for its supplier as well as firms’ industry concentration, I proxy
for pijt as
pijt ≈ % of sales that j represents for i at t× Concentration score in j’s industry at t. (13)
Thus, the higher the percentage customer j represents for supplier i, the higher the likelihood that
shocks affecting j also affect i. In addition, if customer j’s industry is highly concentrated then
supplier i is likely to face problems finding another customer in case j faces distress, other things
being equal. I obtain the percentage of sales that a customer represents for its supplier at a given
year from the Cohen and Frazzini (2008)’s dataset, and industry concentration scores from the
Hoberg and Phillips (2010)’s fitted SIC-based concentration data.
For illustration, figures 8, 9 and 10 depict the time series of cross-sectional distributions of
propensities from 1980 to 1997; namely,pijt(i,j)∈Gn
1997
t=1980. To determine the parameters that
define the dynamics of ζ, I fit a Beta distribution to each cross-sectional distribution of propensities.
The fitted Beta distributions are depicted with dots in figures 8, 9 and 10. From this procedure,
I obtain a time series of estimates for ζ, ζ∗t 2005t=1980, which are depicted in figure 11. I then fit
a vector autoregressive (VAR) process to the time series of estimates ζ∗t 2005t=1980. After doing so,
I discretize the fitted VAR into a four-states Markov chain using Gospodinov and Lkhagvasuren
20The construction of the network emulates a variation of a preferential attachment model and it is similar to astatic version of the model proposed by Atalay et al. (2011).
21A cycle consists of a sequence of firm relationships starting and ending at the same firm, with each pair ofconsecutive firms in the sequence directly connected to each other in the network.
18
(2014)’s method and obtain ζ∗1L = 0.90, ζ∗1H = 1.19, ζ∗2L = 52.41, ζ∗2H = 72.70, and
Ω∗ =
0.61 0.16 0.16 0.04
0.16 0.63 0.04 0.17
0.17 0.04 0.63 0.16
0.04 0.16 0.16 0.61
,
with a stationary distribution given by P(ζ∗LH) = P(ζ∗HL) = P(ζ∗LL) = P(ζ∗HH) = 0.25.
Figures 12, 13 and 14 depict the centrality of a relationship as a function of its propensity to
transmit shocks for each customer–supplier network from 1980 to 1997. As these figures suggest,
there is a slight negative relationship between the propensity of relationships to transmit shocks and
relationships centrality, as the relationships of peripheral firms tend to exhibit higher propensities
than the relationships of central firms.
It is important to note one important caveat regarding the selection of the network topology
using this database. Because firms need to be sufficiently large to be publicly traded and to
represent at least 10% of the annual sales of a publicly traded company, many U.S. firms and their
relationships are overlooked. As a consequence, one may be able to construct, in the most favorable
case, a network that resembles a small part of the aggregate U.S. economy. To partially ensure that
the topology of the benchmark economy provides a fair representation of the network that underlies
the aggregate U.S. economy, I compare the topology of the benchmark economy with the topology
of networks constructed from BEA input–output tables. As table VII shows, the network in the
benchmark economy does a good job representing some features of the U.S. input–output network
and, in doing so, potentially provides a reasonable representation of the aggregate U.S. economy.22
22It is an empirical issue whether a network uncovered using BEA input–output tables provides a sensible repre-sentation of the network structure that underlies the U.S. economy—I leave this for future research. Another way touncover the underlying network using the framework in this paper is to use probabilistic graphical models—which arecommonly used to represent statistical relationships in large and complex systems—as my baseline model predictscertain behavior of return covariances across stocks. For instance, one may calibrate the network using a graphicallasso estimator (GLASSO) to match observed return covariances. In doing so, one estimates an undirected and tem-porally invariant network by estimating a sparse inverse covariance matrix using a lasso (L1) penalty as in Friedman,Hastie, and Tibshirani (2008). The basic estimation strategy assumes that observations have a multivariate Gaussiandistribution with mean µ and covariance matrix Σ. If the ijth component of Σ−1 is zero, then variables i and jare conditionally independent, given the rest of the variables, which is graphically represented as the lack of an edgebetween variables i and j in Gn. The normality assumption can be relaxed as in Liu et al. (2012).
19
B. Selecting the rest of the parameter values
For the sake of illustration, the rest of the parameters can be separated into three groups. Table
IV reports the key parameter values in the calibrated model. Parameters in the first group define
the preferences of the representative investor, which I select in line with Bansal and Yaron (2004).
Thus, β = 0.997, γ = 10 and ρ = 0.65 (IES ≈ 1.5).
Parameters in the second group define the dynamics of firms’ cash flows. I use annual data
on earnings per share from COMPUSTAT to proxy for firms’ cash flows. I restrict my focus to
firms mentioned in the customer–supplier database, as the number of direct relationships of a
firm is known only for such firms. To determine parameters α0 and α1, I run cross-sectional OLS
regressions specified by equation (1) at an annual frequency. To run such cross-sectional regressions,
I need to determine whether firm i faces a negative cash-flow shock in any given year. To do so,
I exploit the temporal variation of firms’ cash flows and run time series regressions at the firm
level, correcting for the existence of time trends. In particular, I run the following n time series
regressions,
log
(yi,tYt−1
)= β0 + β1 ∗ t+ ǫt, (14)
and consider that firm i faces a negative shock at year t if log(
yi,tYt−1
)is below the value predicted by
equation (14) for more than one standard deviation of the residuals computed from equation (14).
This procedure allows me to potentially identify the years in which firms faced negative cash-
flow shocks and compute annual estimates for α0 and α1, which are depicted in figure 16. I set
α0 = 0.27 and α1 = 0.05, which correspond to the averages of annual estimates.23 For simplicity,
I set α2 = 0.0626 and q = 0.129 so that the unconditional mean and volatility of dividend growth
generated by the calibrated model are similar to the ones found in the data.24 Appendix C describes
the method used to simulate the model.
Parameters in the third group define the difference between aggregate output and consumption
23Estimates of α0 and α1 are statistically significant for most of the years in the sample.24To determine the benchmark values of α0 and α1 at a monthly frequency, I assume that yi,year = 12× yi,month,
with i ∈ 1, · · · , n. Provided that the data on firms’ cash flows are at an annual frequency, this assumptionfacilitates the computation of parameters α0 and α1 at a monthly frequency because Yyear = 12 × Ymonth so that
log(
yi,year+1
Yyear
)= log
(yi,month+1
Ymonth
).
20
growth. Within the baseline model, output growth equals consumption growth at equilibrium. To
provide a more realistic description of dividends and improve the fit of the calibrated model to
data, I augment the baseline model so that consumption and dividends are two different processes.
Similar to many others, including Cecchetti, Lam, and Mark (1993), Abel (1999), Campbell (1999,
2003), and Bansal and Yaron (2004), I assume that dividend and consumption growth jointly satisfy,
xt+1 = µ+ τ∆ct+1 + σxξt+1. (15)
Parameters µ and τ are constant and ξt+1 is an i.i.d. normal with zero mean and unit variance. Thus,
the representative investor is implicitly assumed to have access to labor income in the augmented
model. For simplicity, ξt+1 is independent of both ∆ct+1 and variables εi,t+1ni=1. As in Abel
(1999), parameter τ represents the leverage ratio on equity. If µ = σx = 0 and τ = 1, then the
market portfolio is a claim to total wealth and the baseline model is recovered. I follow Bansal
and Yaron (2004) and set τ = 3. I set µ = −0.019/12 so that the difference between unconditional
means of consumption and dividend growth generated by the calibrated model is similar to the
one found in the data. Finally, I set σx = 0.05/√12 so that the difference between unconditional
volatilities of dividend and consumption growth generated by the calibrated model is similar to the
one found in the data.25
V. Implications of the Calibrated Model
This section quantitatively evaluates the ability of the calibrated model to rationalize dynamics
of stock returns. It shows that changes in the propagation of shocks within production networks that
resemble U.S. customer–supplier networks are quantitatively important to understanding variations
in stock returns in both the aggregate and the cross section. Section V.A shows that the model
generates long-run risks in consumption, high and volatile risk premia and a low and stable risk-
free rate. Section V.B shows that the model generates a realistic return spread between central
and peripheral firms in the network, which cannot be accounted for by standard asset pricing
25 Despite the fact that aggregate output and consumption are two different processes within the augmentedmodel, both of these processes are still determined by the propagation of shocks within the network economy. Inparticular, the distribution of xt+1 is fully determined by the propagation of shocks, as equation (3) shows, whereasthe distribution of ∆ct+1 is also determined by the propagation of shocks, as equation (15) states.
21
models. Section V.C shows that the model also generates a high degree of common time variation
in firm-level return volatilities, which is aligned with recent empirical evidence.
A. Customer–Supplier Networks and Long-Run Risks
Table V exhibits moments generated under the benchmark parameterization. By construction,
the benchmark parameterization delivers annual averages and volatilities of consumption and divi-
dend growth similar to those found in the data. It also delivers an average market return of 12%,
an annual volatility of the market return of 18.92%, an average risk-free rate of 2.16%, an annual
volatility of the risk-free rate of 0.7%, an annual equity premium of 10%, and an average Sharpe
ratio of 0.52. With the exception of the volatility of the risk-free rate and Sharpe ratio, all values
are aligned with those found in the data.
Besides matching the above moments, the calibrated model generates a persistent component
in expected consumption growth and stochastic consumption volatility similar to those assumed
by the long-run risks (LRR) model of Bansal and Yaron (2004). As Bansal and Yaron (2004)
and Bansal, Kiku, and Yaron (2012) show, these two features, together with Epstein-Zin-Weil
preferences, help to quantitatively explain an array of important asset market phenomena.26 Table
VI reports summary statistics of several similarity measures of time series generated with either
the calibrated model or the LRR model. To compute averages and standard deviations of these
similarity measures, I sample from the calibrated model and the LRR model to construct two
distributions for each similarity measure: one for expected consumption growth, Et [∆ct+1], and
one for the conditional volatility of consumption growth, Volt [∆ct+1]. Reported values are based
on 300 simulated economies over 620 periods. The first 100 periods are disregarded to eliminate
bias coming from the initial condition. As table VI suggests, both models generate similar time
series for conditional expected consumption growth and conditional consumption volatility.
It is important to appreciate that the persistent component in expected consumption growth
and stochastic consumption volatility are endogenously generated within my model rather than
26Since Bansal and Yaron (2004), several authors have used the long-run risk framework to explain an array ofmarket phenomena. For instance, Kiku (2006) provides an explanation of the value premium within the long-runrisks framework. Drechsler and Yaron (2011) show that a calibrated long-run risks model generates a variancepremium with time variation and return predictability that is consistent with the data. Bansal and Shaliastovich(2013) develop a long-run risks model that accounts for bond return predictability and violations of uncovered interestparity in currency markets.
22
exogenously imposed, as in many asset pricing models. The calibrated model generates these
two features because of two reasons: (1) inter-firm relationships are long-term, and (2) the four
parameter vectors estimated from the data, ζ∗LL, ζ∗LH , ζ∗HL, and ζ∗HH , generate similar cross-sectional
distributions of pijt(i,j)∈Gn, as figure 15 shows. As a consequence, the propagation mechanism of
shocks across firms is fairly stable over time.
While the model endogeneously generates long-run risk, it does not provides a complete micro-
foundation of long-run risks because inter-firm relationships are exogenously determined. Nonethe-
less, the model provides a novel link between equilibrium asset prices and the propagation of firm
level shocks within customer–supplier networks that is consistent with the existence of long-run
risks. In doing so, the model provides a new perspective on the potential sources of long-run risks
and suggests that small changes in the propagation mechanism of shocks in fairly sticky production
networks are quantitatively relevant to understanding asset market phenomena.
B. Firms’ Centrality and the Cross Section of Risk Premiums
Besides helping to explain aggregate asset market phenomena, the model helps to understand
the cross section of expected returns because it provides a mapping between firms’ quantities of
priced risk and firms’ importance in the network. To measure the importance of a firm in the
network, I define the centrality of firm i at period t as the average number of firms that can be
affected by a shock to firm i at period t. This measure captures the relative importance of firm i
in transmitting shocks to other firms in the economy at period t. Provided that the cross-sectional
distribution of pij(i,j)∈Gnchanges over time, firms’ centrality scores also change over time.
To quantitatively assess the effect of a firm’s importance in the network on a firm’s return–
risk trade off, I simulate the benchmark economy at a monthly frequency and construct portfolios
based on centrality. Firms are assigned into centrality terciles once per year, and the value-weighted
portfolios are not rebalanced for the next 12 months. Using simulated data, a portfolio that is long
in the low centrality tercile portfolio and short the high centrality tercile portfolio generates a
statistically significant annual return of 3.8% (0.32% per month). Such a return is computed using
200 simulated economies over 1100 monthly observations. I disregard the first 100 observations in
each simulation to eliminate the potential bias coming from the initial condition.
The above result is explained by the fact that relationships of peripheral firms tend to exhibit
23
higher propensities to transmit shocks than relationships of central firms. As a consequence, pe-
ripheral firms tend to have higher exposure to negative shocks that affect their partners. Such a
contagion risk overweights the potential benefits that peripheral firms receive from their few rela-
tionships and, thus, peripheral firms command higher risk premiums than central firms on average.
However, central firms benefit from the diversification of their customers and suppliers because their
relationships exhibit, on average, small propensities to transmit shocks and, thus, their contagion
risk is overweighted by the benefits generated by their many relationships.
Table VIII shows that the calibrated model generates a realistic spread between low and high
centrality portfolios as the average annual return difference between low and high centrality port-
folios in the Cohen and Frazzini (2008) database is 4.1% (0.34% per month). Table VIII reports
monthly average returns, alphas—from the CAPM, Fama and French (1993) three–factor model,
Carhart (1997) four–factor model, and Fama and French (2015) five–factor model—and loadings
from the five–factor model of Fama and French (2015) for three portfolios of stocks sorted by annual
centrality as well as the portfolio that is long the low centrality tercile and short the high central-
ity tercile. Firms are assigned into centrality terciles at the end of October every year and the
value-weighted portfolios are not rebalanced for the next 12 months. The sample is from June 1981
to December 2004. The last five columns report betas for the five–factor model of each centrality
tercile.
As table VIII suggests, there is a strong negative relation in the data between firms’ centrality
and future returns that cannot be captured by commonly used asset pricing models. Firms in
the low centrality tercile command an average monthly return of 2.06%, whereas firms in the high
centrality tercile command an average monthly return of 1.71%. The 0.34% monthly difference
in returns between these two portfolios is economically and statistically significant and appears
naturally in an equilibrium context, like the one illustrated in this paper, as a compensation for
contagion risk.
To better understand the sources of the difference in future returns between firms with different
centrality scores, I focus on manufacturing and service firms because they jointly represent almost
70% of the firms in the dataset. Table IX considers only manufacturing firms, table X considers
only service firms, and table XI considers both manufacturing and service firms. As in table
VIII, tables IX, X, and XI report raw returns, alphas, and betas of each centrality tercile and the
24
portfolio that is long the low centrality tercile and short the high centrality tercile. They also
report alphas from the CAPM, and the three–, four– and five–factor models. Consistent with
empirical evidence documented by Wu and Birge (2014), these tables suggest that central firms in
manufacturing and service industries may benefit from the diversification of customers and suppliers.
Such diversification helps central firms to mitigate contagion risk better than less central firms and,
thus, command lower risk premiums.
C. Factor Structure on Firm-Level Return Volatility
The calibrated model also generates a high degree of common time variation in return volatilities
at the firm level, which is aligned with recent empirical evidence, e.g., Herskovic et al. (2014) and,
Duarte et al. (2014). To facilitate comparison with evidence documented by Herskovic et al. (2014),
figure 17 illustrates the annual total return volatility at the firm-level averaged within start-of-year
size quintiles. As figure 17 shows, firms of all sizes exhibit similar time series volatility patterns
in the calibrated model. On average, the first principal component of the cross section of annual
return volatility accounts for 99% of the variance. Within the model, the existence of this factor
structure is not surprising, as fluctuations in the cross-sectional distribution of pijt(i,j) ∈ Gndrive
changes in growth opportunities and uncertainty across firms, which translate into changes in prices
and returns at the firm-level. Provided that stock returns respond to a common factor—given by
the state of ζ—firm-level return volatilities inherit a factor structure.27
VI. Conclusion
This paper develops a dynamic equilibrium model to study the asset pricing properties that
stem from the propagation of shocks across firms in a network economy. The fundamental insight
is that extending standard asset pricing models to take into account the way that shocks propagate
within fairly sticky production networks can make significant progress toward generating a unifying
27Recent empirical evidence also suggests the existence of common time variation in firm-level idiosyncratic volatil-ities, e.g., Herskovic et al. (2014) and Duarte et al. (2014). In unreported results, I explore the extent to whichidiosyncratic volatilities exhibit a factor structure within the calibrated model. After removing the market as a com-mon factor of return volatilities, the high degree of common time variation in firm-level return volatilities tends todisappear. On average, the first principal component of the cross section of annual idiosyncratic volatility accountsonly for 3% of the variance (see figure 17(b)).
25
framework that simultaneously captures dynamics of the aggregate and the cross section of stock
returns.
A calibrated model that matches key features of customer–supplier networks in the United
States generates long-run risks, high and volatile risk premiums, and a low and stable risk-free rate.
In the model, low-frequency changes in the shock propagation mechanism endogenously generate
persistence in firms’ growth prospects, which, in turn, drives a small and persistent component
in expected aggregate consumption growth. With recursive preferences, sizeable risk premiums
arise because investors fear that extended periods of low economic growth coincide with low asset
prices. Similarly, a small risk-free rate is driven by investors saving for long periods of low economic
growth.
The model also helps in understanding the cross section of expected returns, as it provides a
mapping between firms’ quantities of priced risk and firms’ importance in the network. In the
calibrated economy, firms that are more central in the network command lower risk premiums than
firms that are less central: central firms tend to benefit from the diversification of their customers
and suppliers and, thus, they mitigate contagion risk better than less central firms. In the time
series, firm-level return volatilities exhibit a high degree of co-movement. These two predictions
are consistent with data from firms in manufacturing and service industries and recent empirical
evidence but cannot be accounted for by standard asset pricing models.
REFERENCES
Abel, Andrew B. 1999. “Risk premia and term premia in general equilibrium.” Journal of Monetary
Economics 43:3–33.
Acemoglu, Daron, Vasco M. Carvalho, Asuman Ozdaglar, and Alireza Tahbaz-Salehi. 2012. “The
Network Origins of aggregate fluctuations.” Econometrica 80:1977–2016.
Acemoglu, Daron, Asuman Ozdaglar, and Alireza Tahbaz-Salehi. 2015. “Microeconomic Origins of
Macroeconomic Tail Risks.” NBER Working Paper (20865).
Ahern, Kenneth R. 2013. “Network Centrality and The Cross-Section of Stock Returns.” USC -
Marshall School of Business Working Paper .
26
Atalay, Enghin, Ali Hortacsu, James Roberts, and Chad Syverson. 2011. “Network Structure of
Production.” Proceedings of the National Academy of Sciences of the United States of America
108:5199–5202.
Bansal, Ravi, Dana Kiku, and Amir Yaron. 2012. “An Empirical Evaluation of the Long Run Risks
Model for Asset Prices.” Critical Finance Review 1:183–221.
Bansal, Ravi and Ivan Shaliastovich. 2013. “A Long-run Risks Explanation of Predictability Puzzles
in Bond and Currency Markets.” Review of Financial Studies 26:1–33.
Bansal, Ravi and Amir Yaron. 2004. “Risks for the Long Run: A Potential Resolution of Asset
Pricing Puzzles.” Journal of Finance 59:1481–1509.
Barrot, Jean-Noel and Julien Sauvagnat. 2016. “Input Specificity and the Propagation of Idiosyn-
cratic Shocks in Production Networks.” Quarterly Journal of Economics 131:1543–1594.
Bidder, Rhys and Ian Dew-Becker. 2016. “Long-Run Risk Is the Worst-Case Scenario.” American
Economic Review 106:2494–2527.
Billingsley, Patrick. 1995. Probability and Measure. John Wiley and Sons, Inc., third ed.
Blume, Lawrence, David Easley, Jon Kleinberg, Robert Kleinberg, and Eva Tardos. 2013. “Network
Formation in the Presence of Contagious Risk.” Journal ACM Transactions on Economics and
Computation 1:6:1–6:20.
Boehm, Christoph, Aaron Flaaen, and Nitya Pandalai-Nayar. 2015. “Input Linkages and the
Transmission of Shocks: Firm-Level Evidence from the 2011 Tohoku Earthquake.” Working
Paper .
Boone, Audra and Vladimir Ivanov. 2012. “Bankruptcy Spillover Effects on Strategic Alliance
Partners.” Journal of Financial Economics 103:551–569.
Boyarchenko, Nina and Anna Costello. 2015. “Counterparty Risk in Material Supply Contracts.”
Federal Reserve Bank of New York Staff Report 694.
Buraschi, Andrea and Paolo Porchia. 2012. “Dynamic Networks and Asset Pricing.” Working
Paper .
27
Campbell, John. 1999. “Asset prices, consumption and the business cycle.” In Handbook of Macroe-
conomics, edited by John Taylor and Michael Woodford. Elvesier, North-Holland, 3967–4056.
Campbell, John Y. 2003. “Consumption-based asset pricing.” In Handbook of the Economics of
Finance, edited by G.M. Constantinides, M. Harris, and R. Stulz. Elsevier Science, 803–887.
Carhart, Mark. 1997. “On Persistence in Mutual Fund Performance.” Journal of Finance 52:57–82.
Carvalho, Vasco and Xavier Gabaix. 2013. “The Great Diversification and its Undoing.” American
Economic Review 103:1697–1727.
Carvalho, Vasco, Makoto Nirei, and Yukiko Saito. 2014. “Supply Chain Disruptions: Evidence
from the Great East Japan Earthquake.” Working Paper .
Carvalho, Vasco M. 2010. “Aggregate Fluctuations and the Network Structure of Intersectoral
Trade.” Unpublished Manuscript .
Cecchetti, Stephen, Poksang Lam, and Nelson Mark. 1993. “The equity premium and the risk-free
rate.” Journal of Monetary Economics 31:21–45.
Chaney, Thomas. 2014. “The Network Structure of International Trade.” American Economic
Review 104:3600–3634.
———. 2016. “The Gravity Equation in International Trade: An Explanation.” Journal of Political
Economy .
Cohen, Lauren and Andrea Frazzini. 2008. “Economic Links and Predictable Returns.” Journal of
Finance 33:1977–2011.
Colbourn, Charles J. 1987. The Combinatorics of Network Reliability. Oxford University Press.
Demange, Gabrielle and Myrna Wooders. 2005. Group Formation in Economics: Networks, Clubs,
and Coalitions. Cambridge University Press.
Di Giovanni, Julian, Andrei A. Levchenko, and Isabelle Mejean. 2014. “Firms, Des-
tinations, and Aggregate Fluctuations.” Econometrica 82 (4):1303–1340. URL
http://dx.doi.org/10.3982/ECTA11041.
28
Drechsler, Itamar and Amir Yaron. 2011. “What’s Vol Got to Do with It.” Review of Financial
Studies 24:1–45.
Duarte, Jefferson, Avraham Kamara, Stephen Siegel, and Celine Sun. 2014. “The Systemic Risk
of Idiosyncratic Volatility.” University of Washington Working Paper .
Elliott, Matthew, Benjamin Golub, and Matthew Jackson. 2014. “Financial Networks and Conta-
gion.” American Economic Review 104:3115–3153.
Fama, Eugene and Kenneth French. 2015. “A five-factor asset pricing model.” Journal of Financial
Economics 116:1–22.
Fama, Eugene F. and Kenneth R. French. 1993. “Common risk factors in the returns on stocks
and bonds.” Journal of Financial Economics 33:3–56.
Friedman, Jerome, Trevor Hastie, and Robert Tibshirani. 2008. “Sparse inverse covariance estima-
tion with the graphical lasso.” Biostatistics 9:432–441.
Gabaix, Xavier. 2011. “The Granular Origins of Aggregate Fluctuations.” Econometrica 79:733–
772.
Gamarnik, David. 2013. “Correlation Decay Method for Decision, Optimization, and Inference in
Large-Scale Networks.” In Theory Driven by Influential Applications. 108–121.
Garcia, Rene, Eric Renault, and A. Semenov. 2006. “Disentangling Risk Aversion and Intertemporal
Substitution.” Finance Research Letters 3:181–193.
Gospodinov, Nikolay and Damba Lkhagvasuren. 2014. “A Moment-Matching Method for Approx-
imating Vector Autoregressive Processes by Finite-State Markov Chains.” Journal of Applied
Econometrics 29:843–859.
Goyal, Sanjeev. 2007. Connections: An Introduction to the Economics of Networks. Princeton
University Press.
Goyal, Sanjeev and Jose Moraga-Gonzalez. 2001. “R&D Networks.” RAND Journal of Economics
32:686–707.
29
Grimmett, Geoffrey. 1989. Percolation. Springer-Verlag.
Hamilton, James D. 1989. “A New Approach to the Economic Analysis of Nonstationary Time
Series.” Econometrica 57:357–384.
Hansen, Lars, John C. Heaton, J. Lee, and Nicholas Roussanov. 2007. “Intertemporal Substitution
and Risk Aversion.” In Handbook of the Econometrics, edited by J.J. Heckman and E.E. Leamer.
North-Holland, 3967–4056.
Herskovic, Bernard. 2015. “Networks in Production: Asset Pricing Implications.” Unpublished
Manuscript .
Herskovic, Bernard, Bryan Kelly, Hanno Lustig, and Stijn Van Nieuwerburgh. 2014. “The Com-
mon Factor in Idiosyncratic Volatility: Quantitative Asset Pring Implications.” Chicago Booth
Working Paper .
Hertzel, Michael G., Micah S. Officer, Zhi Li, and Kimberly Rodgers Cornaggia. 2008. “Inter-firm
Linkages and the Wealth Effects of Financial Distress along the Supply Chain.” Journal of
Financial Economics 87:374–387.
Hoberg, Gerard and Gordon Phillips. 2010. “Real and Financial Industry Booms and Busts.”
Journal of Finance 65:45–86.
Jackson, Matthew O. 2008. Social And Economic Networks. Princeton University Press.
Jorion, Philippe and Gaiyan Zhang. 2009. “Credit Contagion from Counterparty Risk.” Journal
of Finance 64:2053–2087.
Kalpakis, Konstantinos, Dhiral Gada, and Vasundhara Puttagunta. 2001. “Distance Measures for
Effective Clustering of ARIMA Time-Series.” Proceedings 2001 IEEE International Conference
on Data Mining .
Kaltenbrunner, Georg and Lars A. Lochstoer. 2010. “Long-Run Risk through Consumption Smooth-
ing.” Review of Financial Studies 23:3190–3224.
Kandel, Shmuel and Robert F. Stambaugh. 1991. “Asset returns and intertemporal preferences.”
Journal of Monetary Economics 27:39–71.
30
Karrer, Brian, M.E.J. Newman, and Lenka Zdeborova. 2014. “Percolation on Sparse Networks.”
Physical Review Letters 113.
Kiku, Dana. 2006. “Is the Value Premium a Puzzle?” Unpublished Manuscript .
Kung, Howard and Lukas Schmid. 2015. “Innovation, Growth, and Asset Prices.” Journal of
Finance 70:1001–1037.
Lim, Kevin. 2016. “Firm–to–firm Trade in Sticky Production Networks.” Unpublished Manuscript
.
Liu, Han, Fang Han, John Lafferty, and Larry Wasserman. 2012. “High Dimensional Semiparamet-
ric Gaussian Copula Graphical Models.” Annals of Statistics 40:2293–2326.
Lucas, Robert E. 1978. “Asset Prices in an Exchange Economy.” Econometrica 46:1429–1445.
Lyons, Russell. 1990. “Random Walks and Percolation on Trees.” Annals of Probability 18:931–958.
Mehra, Rajnish and Edward C. Prescott. 1985. “The Equity Premium: A Puzzle.” Journal of
Monetary Economics 15:145–161.
Montero, Pablo and Jose A. Vilar. 2014. “TSclust: An R Package for Time Series Clustering.”
Journal of Statistical Software 62 (1):1–43. URL http://www.jstatsoft.org/v62/i01/.
Newman, M.E.J. 2010. Networks: An Introduction. Oxford University Press.
Oberfield, Ezra. 2013. “Business Networks, Production Chains, and Productivity: A Theory of
Input-Output Architecture.” Unpublished Manuscript .
Stauffer, Dietrich and Amnon Aharony. 1994. Introduction to Percolation Theory. Taylor and
Francis, second ed.
Todo, Yasuyuki, Kentaro Nakajima, and Petr Matous. 2015. “How do Supply Chain Networks affect
the Resilience of Firms to Natural Disasters? Evidence from the Great East Japan Earthquake.”
Journal of Regional Science 55 (2):209–229. URL http://dx.doi.org/10.1111/jors.12119.
Weil, Philippe. 1989. “The Equity Premium Puzzle and the Risk-Free Rate Puzzle.” Journal of
Monetary Economics 24:401–421.
31
Williamson, Oliver E. 1979. “Transaction-Cost Economics: The Governance of Contractual Rela-
tions.” Journal of Law and Economics 22:233–261.
———. 1983. Markets and Hierarchies: Analysis and Antitrust Implications. New York: Free
Press.
Wu, Jing and John R. Birge. 2014. “Supply Chain Network Structure and Firm Returns.” Unpub-
lished Manuscript .
Appendix A: Asymptotic Normality of Consumption Growth
To fix notation, let Gn+1 denote the network Gn, to which I add one new firm and all the relationships the
entrant firm may create with incumbent firms in Gn. The following proposition imposes sufficient conditions on: (a)
the sequence of production networks Gnn≥1 and (b) the values within the propensity matrix, pt+1, so that ∆ct+1
is normally distributed as n grows large.
PROPOSITION 5 (Asymptotic Normality of ∆ct+1): Given 0 < q < 1 and a sequence of production networks,
Gnn≥1, define the threshold probability pqc and the set Cn as
pqc ≡ supp∈(0,1)
p : If every relationship in Gn has propensity p, then lim
n→∞Pαq (Gn) = 0
Cn ≡ G is a connected component of Gn : Number of nodes in G = O(n)
where Pαq (Gn) denotes the probability that a shock to any given firm in Gn affects at least αn firms via shock propagation,
with α > 0. The graph G is said to be a connected component of Gn if G is a subset of Gn in which any two firms are
connected to each other by sequences of relationships and which is connected to no additional firms in Gn. Notation
x = O(n) indicates that x grows, at most linearly, with n. If
limn→∞
max
(i,j) ∈ Cn
pij,t+1
< pqc,
then√nWn,t+1 and ∆ct+1 are normally distributed as n grows large.
Under the conditions of Proposition 5, the distribution of ∆ct+1 can be characterized in terms of its mean and
variance. Because the network topology is fixed, the dynamics of the mean and variance of consumption growth are
fully determined by the dynamics of the propensity matrix pt+1. Provided that the dynamic of pt+1 is determined
by ζt+1, the economy follows a Markov process with a continuum of values for aggregate consumption and its growth
rate, ∆ct+1, but only four values for the first two moments of the distribution of consumption growth.
The following corollaries provide a more detailed characterization of those large network economies in which
consumption growth is normally distributed. Corollary 1 focuses on large networks in which all firms have the same
number of direct relationships.
32
COROLLARY 1 (Symmetric Production Networks): Suppose pij,t+1 = pt+1 > 0, ∀(i, j) ∈ Gn. Given a sequence of
production networks, Gnn≥1, with limiting topology G∞,28
• pqc = 1− 2 sin(
π18
)≈ 0.65 if G∞ is the two-dimensional honeycomb lattice.
• pqc = 12if G∞ is the two-dimensional square lattice.
• pqc = 2 sin(
π18
)≈ 0.34 if G∞ is the two-dimensional triangular lattice.
• pqc = 1z−1
if G∞ is the Bethe lattice with z neighbors per each firm.
Figure 3 illustrates each of the network economies considered in corollary 1. Corollary 2 focuses on large networks
in which the number of direct relationships differs across firms.
COROLLARY 2 (Asymmetric Production Networks): Suppose pij,t+1 = pt+1 > 0, ∀(i, j) ∈ Gn. Given a sequence of
production networks, Gnn≥1,
• pqc = 1branching number of G∞
if G∞ is a tree. The branching number of a tree is the average number of relationships
per firm in a tree.29
• pqc = 1eM
if Gn is sparse and locally treelike. Gn is said to be sparse if the number of relationships in Gn
increases linearly with n, as n increases. Gn is said to be locally treelike if an arbitrarily large neighborhood
around any given firm takes the form of a tree. For finite n, parameter eM is the leading eigenvalue of the
matrix
Mn =
An In −Dn
In 0
(1)
where An is the adjacency matrix of Gn, i.e. the n× n matrix in which Aij = 1 if there is a direct relationship
between firms i and j and zero otherwise. In is the n× n identity matrix, and Dn is the diagonal matrix that
contains the number of relationships per firm along the diagonal.
Appendix B: Proofs
This section contains the proofs of the propositions and corollaries in the body of the paper and Appendix A.
The following computations consider three assumptions:
ASSUMPTION 1: Firm i’s output at period t+ 1, yi,t+1, follows
log
(yi,t+1
Yt
)≡ α0 + α1di − α2
√nεi,t+1 , (1)
where Yt denotes the aggregate output of the economy at t and εi,t+1 denotes a Bernoulli random variable that equals
28A lattice is a graph whose drawing can be embedded in Rn. The two-dimensional honeycomb lattice is a graph
in 2D that resembles a honeycomb. The two-dimensional square lattice is a graph that resembles the Z2 grid. The
two-dimensional triangular lattice is a graph in 2D in which each node has six neighbors.29A tree is a network in which any two nodes are connected by exactly one sequence of edges. A forest is a network
whose components are trees.
33
one if firm i faces a negative shock at t+1 and zero otherwise. Parameter di denotes the number of direct relationships
of firm i. Parameters α0, α1 and α2 are non-negative real numbers.
ASSUMPTION 2: The aggregate output of the economy at t, Yt, is defined as
Yt ≡n∏
i=1
y1/ni,t . (2)
ASSUMPTION 3: Let xt+1 ≡ log(
Yt+1
Yt
)be the log output growth rate of the economy at t + 1 and let ∆ct+1 ≡
log(
Ct+1
Ct
)be the log aggregate consumption growth rate at t+ 1. The processes for xt+1 and ∆ct+1 jointly satisfy
xt+1 = µ+ τ∆ct+1 + σxξt+1, (3)
where µ and τ are constants, σx > 0, and ξt+1d−→ i.i.d. N (0, 1). Variable ξt+1 is independent of ∆ct+1 and all the
variables in the sequence εi,t+1ni=1.
Let st denote the state of the parameter vector ζt. Because the network topology is fixed, st determines the
distributions of aggregate output and consumption growth at t. Provided that ζt follows a Markov process, the
distributions of aggregate output and consumption growth vary over time and the dynamics of the moments of these
distributions satisfy the Markov property.
Sketch of proof of Proposition 5 and Corollaries 1 and 2. Given a sequence of network topologies Gnn≥1 and the
realization of the matrix pt, the goal is to find the conditions under which√nWn,t is normally distributed as n grows
large. Without loss of generality, fix period t so that subscript t on the sequence εi,tni=1 and on the matrix pt can
be eliminated. If the sequence of Bernoulli random variables εini=1 is independent, the Lindeberg-Levy central limit
theorem implies that√nWn is normally distributed as n grows large. Consequently, if p is a matrix of zeros then
εini=1 is a sequence of independent random variables and√nWn is asymptotically normally distributed. In the
presence of inter-firm relationships, however, some elements in the matrix p are different from zero. In particular, εi
and εj are correlated if there exists a sequence of relationships between firms i and j in Gn. In this case, εini=1 is
a sequence of dependent random variables and the conditions under which the Lindeberg-Levy central limit theorem
hold are not satisfied.
Despite the fact that the sequence εini=1 may be dependent,√nWn may still be asymptotically normally
distributed if the dependence among variables in εini=1 is sufficiently weak. In particular, if the sequence εini=1 is
α-mixing and stationary,√nWn follows a normal distribution as n grows large—see Billingsley (1995, Theorem 27.4).
Generally speaking, the sequence εini=1 is said to be α-mixing if εk and εk+n are approximately independent for
large n and k ≥ 1.30 The sequence is said to be stationary if the distribution of the subsequence εl, εl+1, · · · , εl+j
30To be more specific, let αn be a non-negative number such that
∣∣P(A ∩ B)− P(A)P(B)∣∣ ≤ αn (4)
with A ∈ σ(ε1, · · · , εk), B ∈ σ(εk+n, εk+n+1, · · · ), k ≥ 1 and n ≥ 1; where σ(·) denotes the σ-algebra defined on thepower set of 0, 1n ≡ 0, 1 × · · · × 0, 1. The sequence εini=1 is said to be α-mixing if αn → 0 as n grows large.
34
does not depend on the subscript l. A special case of the above result occurs if there exists an ordering of the
sequence εini=1 such that the dependence between variables εk and εj decreases as the distance between them in
such an ordering increases. In particular, if there exists such an ordering and a positive constant m such that the
subsequences ε1, · · · , εk and ε1,k+s, · · · , εk+s+l are independent whenever s > m, the sequence εini=1 is said to
be m-dependent in which case√nWn follows a normal distribution as n grows large.31
In what follows, I apply an idea similar to that behind the notion of m-dependent random variables to sketch
the proof of the asymptotic normality of√nWn. In particular, I impose sufficient conditions on the sequence of
networks Gnn≥1 and on the matrix p so that negative shocks to individual firms tend to remain locally confined as
n grows large. In case the shocks remain locally confined, shocks would almost surely spread only over sets of firms
of finite size—with all sets independent among each other—whose size would become negligible compared to the size
of the economy as n grows large. As a consequence, one would almost surely be able to find an index ordering I
and a positive constant m under which the sequence εii∈I would be m-dependent and, thus,√nWn would follow
a normal distribution as n grows large.
For the sake of illustration, suppose that ∀ (i, j) ∈ Gn, pij = p0 > 0, with i 6= j. Given 0 < q < 1 and a sequence
of networks, Gnn≥1, define the threshold probability pqc as
pqc ≡ supp∈(0,1)
p : lim
n→∞Pαq (Gn) = 0
, (5)
where Pαq (Gn) denotes the probability that a shock to any individual firm in Gn at least affects other αn firms via
shock propagation, with α > 0. Determining the conditions under which a CLT applies to√nWn is related to
determining the threshold probability pqc. If every relationship in Gn has a propensity p0 and p0 < pqc, then shocks
would remain locally confined because the number of firms affected by the shock would become negligible compared
to the size of the economy as n grows large. Then, one would almost surely be able to find an index ordering I and
a positive constant m under which the sequence εii∈I would be m-dependent.
The condition p0 < pqc may be stronger than necessary to prove the asymptotic normality of√nWn. Imposing such
a condition, however, greatly facilitates the proof, as the determination of the threshold pqc has been extensively studied
in percolation theory, e.g. Grimmett (1989) and Stauffer and Aharony (1994). In percolation, pqc is sometimes called
the critical probability or critical phenomenon because it indicates the arrival of an infinite, connected component as
n grows large. A connected component of a graph is a subgraph in which any two nodes are connected to each other
by sequences of edges, and which is connected to no additional nodes in the original graph.
To illustrate how pqc can be determined, consider the following two examples:
• Imagine n firms are arranged in a straight line and each relationship may transmit shocks with probability p0.
The probability that every relationship in the line transmits shocks is pn−10 . Given how shocks spread from
31An independent sequence is 0-dependent using this terminology.
35
one firm to another, the probability that at least one negative shock spreads over n− 1 different firms equals
(1− (1− q)n)P [every relationship in the line transmits shocks] (6)
≈ (1− e−nq)pn−10 (for large n)
which tends toward zero as n→ ∞. Thus, pqc = 1.
• Suppose n firms are arranged in a circle. The probability that every relationships in the circle transmits shocks
is pn0 , which tends toward zero as n→ ∞. Following the previous argument, it is easy to see that pqc = 1.
Taking results from bond percolation, Table I reports critical probabilities for several symmetric network topolo-
gies assuming that every relationship in the graph transmits shocks with the same probability. As Table I shows,
pqc varies across network topologies. For instance, if the limiting topology of the sequence Gnn≥1, G∞, is the two-
dimensional honeycomb lattice, then pqc = 1− 2 sin(
π18
)≈ 0.65, whereas if G∞ is the two-dimensional square lattice
then pqc = 12.
The previous analysis determines conditions under which√nWn is normally distributed for some symmetric net-
work topologies. But what happens in other networks? In particular, under what conditions is√nWn asymptotically
normally distributed in large asymmetric networks? Using random walks on trees, Lyons (1990) shows that if G∞ is
a tree then
pc =1
branching number of G∞, (7)
where the branching number of a tree is the average number of branches per node in a tree.32 A tree is a connected
graph in which two given nodes are connected by exactly one sequence of edges. A tree is said to be z-regular if each
node has degree z. If G∞ is a z-regular tree, the average number of branches per node is z − 1 so pc = 1z−1
; which is
consistent with Table I.33
One can generalize the previous result for topologies where G∞ is sparse and locally treelike. Gn is said to be
sparse if Gn has m edges and m = O(n). Notation m = O(n) indicates that m grows, at most, linearly with n so
there exists a positive number c such that∣∣mn
∣∣ < c for all n. Namely, Gn is sparse if only a small fraction of the
possible n(n−1)2
edges are present. G∞ is said to be locally treelike if in the limit an arbitrarily large neighborhood
around any node takes the form of a tree. Reformulating percolation in trees as a message passing process, Karrer,
32For a concrete definition of the branching number see Lyons (1990, page 935).33To motivate the previous result, it is informative to compute the percolation threshold in the Bethe lattice with
z neighbors per every node. Start at the root and check whether there is a chance of finding an infinite open pathfrom the root. Starting from the root, one has (z − 1) new edges emanating from each new node in each layer of thelattice. Each of these (z − 1) new edges leads to one new node, which is affected with probability p. On average,(z − 1)p nodes are affected at each layer of the lattice. If (z − 1)p < 1 then the average number of affected nodesdecreases in each layer by a factor of (z− 1)p. As a consequence, if (z− 1)p < 1 the probability of finding an infiniteopen path goes to zero exponentially in the path length. Thus, pc = 1
z−1for the Bethe lattice with z neighbors for
every node.
36
Newman, and Zdeborova (2014) show that if G∞ is sparse and locally treelike then
pc =1
ǫH(8)
where ǫH is the leading eigenvalue of the 2n× 2n matrix
M =
A I−D
I 0
(9)
where A is the adjacency matrix that represents Gn, I is the n×n identity matrix, and D is the diagonal matrix with
the number of relationships per firm along the diagonal.
Now to allow for propensities to vary across relationships, one can extend the previous argument in the following
way. Let Cn be the set of graphs composed by those connected components of Gn whose number of nodes grows, at
most, linearly with n. Mathematically,
Cn ≡ G is a connected component of Gn : the number of nodes in G = O(n) (10)
Because in the limit one only needs to pay attention to components of Gn that potentially grow linearly with n, the
following inequality
limn→∞
max
(i,j) ∈ Cn
pij
< pqc, (11)
ensures that shocks to individual firms would remain locally confined provided that sets of firms of finite size become
negligible as the economy grows large. As a consequence, if inequality (11) holds then√nWn and ∆c are normally
distributed as n grows large.
Proof of Proposition 1. I look for an equilibrium such that the price dividend ratio is stationary. I conjecture that
if c is the current aggregate consumption and s the current state of ζt, then Pa(c, s) = was c, in which Pa is the price
of aggregate wealth and was a number that depends on state s. If st = s and st+1 = s′, the realized gross return at
period t+ 1 of the asset that delivers aggregate consumption as its dividend each period, Ra,t+1, equals
Ra,t+1 =Pa,t+1 + Ct+1
Pa,t=
was′ + 1
was
Ct+1
Ct(12)
Setting Ri,t+1 = Ra,t+1 in equation (4) yields,
Et
[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ [
Ra,t+1
] 1−γ1−ρ
= 1
⇒ E
[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ
[wa
s′ + 1
was
Ct+1
Ct
] 1−γ1−ρ
∣∣∣∣s
= 1 (13)
37
Provided that st follows a Markov process, equation (13) can be rewritten as
β1−γ1−ρ
∑
s′=LL,LH,HL,HH
ωs,s′E
((Ct+1
Ct
)1−γ ∣∣∣∣s′)(
was′ + 1
was
) 1−γ1−ρ
= 1 (14)
Reordering equation (14) yields,
was = β
∑
s′=LL,LH,HL,HH
ωs,s′E
(e(1−γ)∆ct+1
∣∣s′)(wa
s′ + 1)1−γ1−ρ
1−ρ1−γ
(15)
which completes the proof.
REMARK 1: If√nWn,t+1 is normally distributed, then
E
(e(1−γ)∆ct+1
∣∣s)
= exp
((1− γ)(α0 + α1d− α2µs − µ)
τ+
(1− γ)2
2
(α22σ
2s − σ2
x
τ 2
))(16)
where
µs ≡ limn→∞
E
(n∑
i=1
εi,t+1√n
∣∣∣∣s)
and σ2s ≡ lim
n→∞Var
(n∑
i=1
εi,t+1√n
∣∣∣∣s)
and the above constants are assumed to be finite so that equation (16) is well-defined.
REMARK 2 (Price of Market Return): If consumption and output growth differ, I compute the price of the market
return as follows. I conjecture that if y is the current aggregate output and s the current state of ζt, then Pm(c, s) =
wms y, where Pm is the price of the market portfolio and wm
s a number that depends on state s. If st = s and st+1 = s′,
then the realized gross return at period t + 1 of the asset that delivers aggregate output as its dividend each period,
Rm,t+1, equals
Rm,t+1 =Pm,t+1 + Yt+1
Pm,t=
wms′ + 1
wms
Yt+1
Yt(17)
Setting Ri,t+1 = Rm,t+1 in equation (4) yields,
Et
[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ [
Ra,t+1
] 1−γ1−ρ
−1
Rm,t+1
= 1
⇒ E
[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ
[wa
s′ + 1
was
(Ct+1
Ct
)] 1−γ1−ρ
−1(wm
s′ + 1
wms
Xt+1
) ∣∣∣∣s
= 1 (18)
where Xt+1 =Yt+1
Yt. Provided that st follows a Markov process, equation (18) can be rewritten as
β1−γ1−ρ
∑
s′=LL,LH,HL,HHωs,s′E
((Ct+1
Ct
)−γ
Xt+1
∣∣∣∣s′)(
was′ + 1
was
) 1−γ1−ρ
−1(wm
s′ + 1
wms
) = 1 (19)
38
Reordering equation (19) yields,
wms = β
1−γ1−ρ
∑
s′∈LL,LH,HL,HHωs,s′E
(e−γ∆ct+1+xt+1
∣∣s′)(wa
s′ + 1
was
) 1−γ1−ρ
−1
(wms′ + 1)
(20)
It follows from (3) that −γ∆ct+1 + xt+1 = µ+ (τ − γ)∆ct+1 + σxξt+1. Therefore, (20) equals to
wms = β
1−γ1−ρ eµ+
σ2x2
∑
s′∈LL,LH,HL,HHωs,s′E
(e(τ−γ)∆ct+1
∣∣s′)(wa
s′ + 1
was
) 1−γ1−ρ
−1
(wms′ + 1)
(21)
Proof of Proposition 2. Setting Ri,t+1 = Rf in equation (4) yields,
E
[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ [
Ra,t+1
] 1−γ1−ρ
−1∣∣∣∣s
=
1
Rf (s). (22)
Provided that st follows a Markov process and Pa(c, s) = was c, the left hand side of equation (22) can be rewritten
as the following sum
β1−γ1−ρ
∑
s′=LL,LH,HL,HH
ωs,s′E
((Ct+1
Ct
)−γ ∣∣∣∣s′)(
was′ + 1
was
) ρ−γ1−ρ
Therefore,
1
Rf (s)= β
1−γ1−ρ
∑
s′=LL,LH,HL,HH
ωs,s′E
(e−γ∆ct+1
∣∣s′)(wa
s′ + 1
was
) ρ−γ1−ρ
,
which completes the proof
Proof of Proposition 3. Consider st = s and st+1 = s′. Equation (4) can be rewritten as,
Pi,t = Et
(Mt+1
(Pi,t+1 + yi,t+1
))i = 1, · · · , n (23)
where
Mt+1 ≡[β
(Ct+1
Ct
)−ρ] 1−γ1−ρ [
Ra,t+1
] 1−γ1−ρ
−1
represents the pricing kernel. Dividing equation (23) by Yt yields
Pi,t
Yt= Et
(Mt+1Xt+1
Pi,t+1
Yt+1
)+ Et
(Mt+1
yi,t+1
Yt
)i = 1, · · · , n (24)
39
which can be rewritten as
vi,t = Et
(Mt+1Xt+1vi,t+1
)+ Et
(Mt+1
yi,t+1
Yt
)i = 1, · · · , n (25)
with vi,t ≡ vi(s) ≡ Pi,t
Yt. Provided that st follows a Markov process and Pa(c, s) = wa
s c, the first term in the right
hand side of equation (25) can be rewritten as
Et
(Mt+1Xt+1vi,t+1
)= β
1−γ1−ρ eµ+
σ2x2
∑
s′=LL,LH,HL,HH
ωs,s′
(wa
s′ + 1
was
) ρ−γ1−ρ
E
(e(τ−γ)∆ct+1
∣∣s′)vi(s
′)
(26)
whereas the second term in the right hand side of equation (25) can be rewritten as
Et
(Mt+1
yi,t+1
Yt
)= eα0+α1diEt
(Mt+1e
−α2√
nεi,t+1
)(27)
The expectation term in the right hand side of equation (27) can be written as
Et
(Mt+1e
−α2√
nεi,t+1
)= β
1−γ1−ρ
∑
s′=LL,LH,HL,HH
ωs,s′E
((Ct+1
Ct
)−γ
e−α2√
nεi,t+1
∣∣∣∣s′)(
was′ + 1
was
) ρ−γ1−ρ
= β1−γ1−ρ
∑
s′=LL,LH,HL,HH
ωs,s′E
(e−γ∆ct+1−α2
√nεi,t+1
∣∣∣∣s′)(
was′ + 1
was
) ρ−γ1−ρ
As a consequence,
vi(s) = β1−γ1−ρ eµ+
σ2x2
∑
s′=LL,LH,HL,HH
ωs,s′
(wa
s′ + 1
was
) ρ−γ1−ρ
E
(e(τ−γ)∆ct+1
∣∣s′)vi(s
′)
+ β1−γ1−ρ eα0+α1di
∑
s′=LL,LH,HL,HH
ωs,s′E
(e−γ∆ct+1−α2
√nεi,t+1
∣∣∣∣s′)(
was′ + 1
was
) ρ−γ1−ρ
i = 1, · · · , n
Define πi(s′) ≡ E [εi,t+1|st+1 = s′]. It is worth noting that
−γ∆ct+1 − α2
√nεi,t+1 = −γ
1
τ
α0 + α1d− α2
∑
j 6=i
εj,t+1√n
− σxξt+1 − µ
︸ ︷︷ ︸
−α2
√n(1− γ
τn
)εi,t+1
= −γ ∆c−i,t+1 − α2
√n(1− γ
τn
)εi,t+1 (28)
Because ∆c−i,t+1 and εi,t+1 are independent
E
(e−γ∆c−i,t+1−α2
√n(1− γ
τn )εi,t+1
∣∣∣∣s)
= E
(e−γ∆c−i,t+1
∣∣s)
πi(s)e−α2
√n(1− γ
τn ) + (1− πi(s))
≈ E
(e−γ∆ct+1
∣∣s)(1− πi(s)) (29)
where the last approximation is accurate for large n. If the distribution of ∆ct+1 is known, the expectation of
40
−γ∆ct+1 − α2√nεi,t+1 can be approximated using equation (29). Therefore,
vi(s) = β1−γ1−ρ eµ+
σ2x2
∑
s′=LL,LH,HL,HH
ωs,s′
(wa
s′ + 1
was
) ρ−γ1−ρ
E
(e(τ−γ)∆ct+1
∣∣ps′)vi(s
′)
+ β1−γ1−ρ eα0+α1di
∑
s′=LL,LH,HL,HH
ωs,s′E
(e−γ∆ct+1
∣∣s)(1− πi(s))
(wa
s′ + 1
was
) ρ−γ1−ρ
i = 1, · · · , n
which completes the proof
Proof of Proposition 4. Recall that
Var(Mt+1
∣∣s)
= E
(M2
t+1
∣∣s)− E
2(Mt+1
∣∣s)
(30)
The first term in the right hand side of equation (30) can be rewritten as
E
(M2
t+1
∣∣s)
= β2(
1−γ1−ρ
) ∑
s′=LL,LH,HL,HH
ωs,s′E
((Ct+1
Ct
)−2γ ∣∣∣∣s′)(
was′ + 1
was
)2(
ρ−γ1−ρ
) (31)
Provided that λm(s) ≡ −Var
(Mt+1
∣∣s)
E
(Mt+1
∣∣s) and E
(Mt+1
∣∣s)= 1
Rf (s), it then follows from equation (30) that
λm(s) =1
Rf (s)−Rf (s)
β2
(1−γ1−ρ
) ∑
s′=LL,LH,HL,HH
ωs,s′
(wa
s′ + 1
was
)2(
ρ−γ1−ρ
)
E
(e−2γ∆ct+1
∣∣s′)
which completes the proof
Appendix C: Simulation of the Model
This section describes the algorithm I use to compute firms’ probabilities of facing negative cash-flow shocks in
each state of nature so one can compute asset prices and returns at the firm level using proposition 4. Let st denote
the state of the parameter vector ζt at period t. To simplify the computation of probabilities πi(st)ni=1, I restrict
the topology of Gn. In general topologies, computing πi(st)ni=1 is difficult, because the number of states that need
to be considered increases exponentially with n. In economies with no cycles, however, computing πi(st)ni=1 is
easier. In those economies, computing πi(st)ni=1 can be framed as a recursive problem, as the following algorithm
describes.
Algorithm Firms’ Probabilities (Gn, st, q)
(∗ Description: Algorithm that computes firms’ probabilities of facing negative shocks if Gn is a forest ∗)
Input: Gn (a forest), st, q.
Output: The set of probabilities of firms facing a negative cash-flow shock at time t, πi(st)ni=1
1. for each firm i ∈ Gn
41
2. Determine the subgraph of Gn wherein firm i participates. Denote such a graph as Ti and label firm i
as its root.34
3. if firm i has a no connections
4. return πi(st) = q
5. else return Prob(i,Ti,st,q)
where Prob(i,Ti,st,q) corresponds to the following recursive program:
Algorithm Prob(i,Ti,st,q)
(∗ Description: Recursive algorithm that computes firm i’s probability of facing a negative cash-flow shock ∗)
Input: A node i in Gn, the tree Ti wherein node i is the root, st, and q.
Output: πi(st)
1. Determine the set of children of node i in Ti, say Ci.35
2. if Ci = ∅
3. return πi(st) = q
4. else if every node in Ci has no children
5. return πi(st) = q + (1− q)(1− E
[∏k∈Ci
(1− qpikt)∣∣st])
6. else return πi(st) = q + (1− q)(1− E
[∏k∈Ci
(1− piktProb(k, Ti,k, st, q))) ∣∣st
]36
In economies with no cycles, it is also simple to compute the first two moments of the distribution of√nWn,t+1
at t+ 1. Let µs, σ2s denote the mean and variance of
√nWn,t+1 if st+1 = s, respectively. In other words,
µs = limn→∞
E
(n∑
i=1
εi,t+1√n
∣∣∣∣s)
= limn→∞
n∑
i=1
πi(s)√n
s = LL,LH,HL,HH. (1)
and
σ2s = lim
n→∞Var
(n∑
i=1
εi,t+1√n
∣∣∣∣s)
(2)
= limn→∞
1
n
n∑
i=1
πi(s) (1− πi(s)) +1
n
∑
(i,j)∈Gn
Cov(εi,t+1, εj,t+1
∣∣st = s) s = LL, LH,HL,HH.
The second term in the equation above can be simplified further. If there exists a path between firms i and j after
edges are removed at period t + 1, then εi,t+1 = εj,t+1. If there is no path between firms i and j in Gn, variables
εi,t+1 and εj,t+1 are independent. It then follows that
Et
[εi,tεj,t
∣∣∣∣ there is a path between i and j
]= Vart [εi,t] + E
2t [εi,t] = πi(s)(1− πi(s)) + π2
i (s)
Et
[εi,tεj,t
∣∣∣∣ there is no path between i and j
]= Et [εi,t]Et [εj,t] = πi(s)πj(s).
34Note that such a graph is a tree provided that Gn is a forest.35In a rooted tree, the parent of a node is the node connected to it on the path to the root. Every node except
the root has a unique parent. A child of a node v is a node of which v is the parent.36Tree Ti,k denotes the branch of tree Ti that starts at node k.
42
Hence,
Et [εi,tεj,t] = Et
[εi,tεj,t
∣∣∣∣ there is a path between i and j
]P [there is a path between i and j at t]
+ Et
[εi,tεj,t
∣∣∣∣ there is no path between i and j
]P [there is no path between i and j at t]
=(πi(s)(1− πi(s)) + π2
i (s))Pij(s) + πi(s)πj(s) (1−Pij(s)) ,
where Pij(s) ≡ P [there is a path between i and j if st = s]. Thus,
Covt [εi,t, εj,t] = Et [εi,tεj,t]− Et [εi,t]Et [εj,t]
=(πi(s)(1− πi(s)) + π2
i (s))Pij(s) + πi(s)πj(s) (1−Pij(s))− πi(s)πj(s)
= πi(s) (1− πj(s))Pij(s).
Therefore,
σ2s = lim
n→∞
1
n
n∑
i=1
πi(s) (1− πi(s)) +1
n
∑
(i,j)∈Gn
πi(s) (1− πj(s))Pij(s)
s = LL,LH,HL,HH.
To compute Pij(s) I need to determine the set of paths that connect firms i and j on Gn. If there is more than
one path connecting firms i and j, computing Pij(s) is difficult because shocks can be transmitted by any of those
paths. On the other hand, if there is only one path connecting any two given firms, say firms i and j, computing
Pij(s) becomes easier because there is a unique path connecting firms i and j. It then becomes handy to restrict the
topology of Gn so that it does not have cycles. The following remark describes Pij(s) when Gn is a forest.
REMARK 3: Suppose Gn is a forest; namely, there are no cycles. Then, every component of Gn is a tree. Provided
that any two given firms are jointed by a unique path (in case such a path exists),
Pij (s) =
E
[∏(k,l)∈Pi,j
pkl
∣∣∣∣s]
where Pi,j is the (unique) path between i and j in Gn
0 there is no path between i and j.
(3)
Appendix D: Tables and Figures
This section contains the tables and figures mentioned in the paper and in the appendix.
43
Table I
Critical probability for different symmetric network topologies
The table reports critical probabilities for different symmetric network topologies. Besides reporting the two examples
described in Appendix A, the table reproduces a subset of the values reported in Stauffer and Aharony (1994, Table
1). The first column reports the topology of G∞. The second column reports the number of neighbors of any given
node in G∞. The third column reports the critical probability, pc(G∞). Despite the fact that G∞ may be highly
connected, if max pt < pc(G∞), then no infinite component emerges as n → ∞, and thus√nWn is asymptotically
normally distributed. For illustrative purposes, figure 3(a) depicts a 2D honeycomb lattice, figure 3(b) depicts a 2D
squared lattice, figure 3(c) depicts a 2D triangular lattice, and figure 3(d) depitcs a Bethe lattice with z = 3. The
Bethe lattice of degree z is defined as an infinite tree in which any node has degree z. For finite n such topologies
are called Cayley Trees.
Topology of G∞ Number of neighbors pc(G∞)
Infinite line (1D lattice) 2 1Infinite circle 2 12D honeycomb lattice 3 1− 2 sin
(π18
)
2D squared lattice 4 12
2D triangular lattice 6 2 sin(
π18
)
Bethe lattice z 1z−1
Table II
Major Industry Groups in Customer–Supplier Database
The table reports the distribution of firms across major industry groups in the Cohen and Frazzini (2008) customer–
supplier database for which monthly return data were available from 1980 to 2004. Major industry groups are defined
by the first two numbers of SIC codes.
Industry Number of firms
Agriculture, forestry, and fishing 14Construction 32Finance, insurance, and real estate 308Manufacturing 2,639Mining 337Public Administration 7Retail 240Service 1,099Transportation, communications, electric, gas, and sanitary 432Wholesale 234
Total 5,342
44
Table III
Characteristics of Customer–Supplier Networks
The table reports characteristics of customer-supplier networks generated at an annual frequency using the Cohen
and Frazzini (2008) dataset from 1980 to 2004. Two firms are connected in the network of year t if one of them
represents at least 10% of the other firm’s sales during that year. The number of components (clusters) in each
network is computed via two consecutive depth-first searches. Provided that degree distributions exhibit fat tails,
one can approximate them via power law distributions in at least the upper tail. Namely, the probability of a given
degree d in the network of year t, Pt(d), can be expressed as P
t (d) = atd−ξt , where at > 0 and ξt > 1 are the
parameters to be estimated. The last row shows the average and standard deviation of the maximum likelihood
estimators for ξt over the sample period. Finally, the column benchmark reports the characteristics of the network
in the benchmark economy.
Characteristic Mean Standard Deviation Benchmark
Number of firms per customer–supplier network 1,038 415 1,038Number of relationships per customer–supplier network 1,066 635 918Number of components per network 174 25 120Size of the largest component 482 382 482Size of the second largest component 54 31 54Size of the third largest component 19 17 19Size of the fourth largest component 11 7 11Size of the fifth largest component 9 5 9Exponent of fitted power law to the degree distribution 2.36 0.18 2.36
Table IV
Benchmark Parameterization
The table reports the list of parameter values in the benchmark parameterization. I set µ = −0.019/12 so the
difference between unconditional means of consumption and dividend growth generated by the benchmark economy
are similar to the ones found in data. I follow Bansal and Yaron (2004), and I set τ = 3. I set σx = 0.05/√12 so
that the difference between unconditional means of consumption and dividend growth generated by the benchmark
economy is similar to the ones found in data. I divide the rest of the parameter values into three groups. Parameters in
the first group define the preferences of the representative investor: β represents the time discount factor, γ represents
the coefficient of relative risk aversion for static gambles, and ρ represents the inverse of the inter-temporal elasticity
of substitution. Parameters in the second group describe firms’ cash-flows: α0 measures the part of firms’ cash-flows
unrelated to inter-firm relationships, α1 measures the marginal benefit a firm receives from each relationship, and α2
measures the decrease in a firm’s cash flow if a firm faces a negative shock. Given a network topology, parameters
in the third group define the stochastic process that determines the propagation of shocks across firms. Parameter q
measures how frequently firms face negative idiosyncratic shocks. The rest of parameters define the cross-sectional
distribution from which propensities of relationships to transmit shocks are determined: ζ1L, ζ1H , ζ2L, and ζ2H .
Preferences Firms’ Cash-flows Propagation of shocksβ γ ρ α0 α1 α2 ζ1L ζ1H q ζ2L ζ2H0.997 10 0.65 0.27 0.05 0.0626 0.90 1.19 0.129 52.4 72.7
45
Table V
Moments under the Benchmark Parameterization
The table reports the first two moments of consumption and dividend growth as well as a set of key asset pricing
moments. Column Data reports moments found in data. Column Model reports moments generated under the
benchmark parameterization described in Table IV. Column BY2004 reports moments generated under the long-run
risks model of Bansal and Yaron (2004). Data on consumption and dividends is obtained from Robert Shiller’s
website http://www.econ.yale.edu/ shiller/data.htm. Moments on the return on aggregate wealth, risk-free rate,
equity premium, and Sharpe ratio are based on data from 1928 to 2014 and obtained from Aswath Damodaran’s
website: http://pages.stern.nyu.edu/∼adamodar/. The annual return on aggregate wealth is approximated by the
annual return of the S&P 500. The return on the risk-free asset is approximated by the yield on three-month T-bills.
All values are in percentage with the exception of average Sharpe ratios.
Moments Data Model BY2004
Average annual log of consumption growth rate 1.9 1.9 1.8Annual volatility of log consumption rate 3.5 3.5 2.8Average annual log dividend growth rate 3.8 3.8 1.8Annual volatility of the log dividend growth rate 11.63 11.7 12.3Average annual market return (S&P 500) 11.53 12 7.2Annual volatility of the market return 19 18.92 19.42Average annual risk-free rate (3-month T-bill) 3.53 2.16 0.86Annual volatility of risk-free rate 3 0.7 0.97Average annual equity risk premium 8 10 6.33Average annual Sharpe ratio 0.4 0.52 0.33
46
Table VI
Similarities between the calibrated model and the LRR model
The table reports averages and standard deviations of similarity measures between time series generated with either
the calibrated model or the benchmark parameterization in the LRR model of Bansal and Yaron (2004). To compute
averages and standard deviations, I sample from the calibrated model and the LRR model to construct two finite-
sample empirical distributions for each similarity measure: one for expected consumption growth, Et [∆ct+1], and
one for the conditional volatility of consumption growth, Volt [∆ct+1]. Reported values are based on 300 simulated
samples over 620 periods. The first 100 periods in each sample are disregarded to eliminate bias coming from the
initial condition. All similarity measures report scores computed as 11+distance
, where distance is defined according to
each similarity measure. Let XT = (X1, · · · , XT ) and YT = (Y1, · · · , YT ) denote realizations from two time series,
X = Xt and Y = Yt. The first and second similarity measures focus on the proximity between X and Y at
specific points of time. The euclidean distance (ED) is defined as√∑T
t=1(Xt − Yt)2, whereas the dynamic time
warping (DTW) distance is defined as minr
(∑mi=1 |Xai − Ybi |
), where r = ((Xa1 , Yb1), · · · , (Xam , Ybm)) is a sequence
of m pairs that preserves the order of observations, i.e., ai < aj and bi < bj if j > i. DTW seeks to find a mapping
such that the distance between X and Y is minimized. This way of computing distance allows two time series
that are similar but locally out of phase to align in a nonlinear manner. The third measure focuses on correlation-
based distances. It uses the partial autocorrelation function (PACF) to define the distance between time series. In
particular, distance is defined as√
(ρXt − ρYt)′Ω(ρXt − ρYt), where Ω is a matrix of weights, whereas ρXt and ρYt
are the estimated partial autocorrelations of X and Y , respectively. The fourth and fifth measures assume that a
specific model generates both time series. The idea is to fit the specific model to each time series and then measure
the dissimilarity between the fitted models. The fourth measure computes the distance between two time series
as the ED between the truncated AR operators. In this case, distance is defined as√∑k
j=1(ej,Xt − ej,Yt)2, where
eXt = (e1,Xt , · · · , ek,Xt) and eYt = (e1,Yt , · · · , ek,Yt) denote the vectors of AR(k) parameter estimators for X and
Y , respectively. The fifth measure computes dissimilarity between two time series in terms of their linear predictive
coding in ARIMA processes, as in Kalpakis, Gada, and Puttagunta (2001). The last measure defines distance based
on nonparametric spectral estimators. Let fXT and fYT denote the spectral densities of XT and YT , respectively. In
this case, the dissimilarity measure is given by a nonparametric statistic that checks the equality of the log-spectra
of the two time series. It defines distance as∑n
k=1
[Zk − µ(λk)− 2 log(1 + eZk−µ(λk))
]−∑n
k=1
[Zk − 2 log(1 + eZk)
],
where Zk = log(IXT (λk)) − log(IYT (λk)), and µ(λk) is the local maximum log-likelihood estimator of µ(λk) =
log(fXT (λk))− log(fYT (λk)) computed with local lineal smoothers of the periodograms. All similarity measures are
computed using the R package TSclust (see Montero and Vilar (2014)).
Et [∆ct+1] Volt [∆ct+1]Similarity measure Mean Standard Deviation Mean Standard Deviation
ED 0.958 0.012 0.974 0.008DTW 0.758 0.091 0.723 0.105PACF 0.736 0.043 0.743 0.043ED in AR 0.908 0.100 0.910 0.097Linear predictive in ARIMA 0.726 0.325 0.729 0.313Spectral distance 1.0 0.000 1.0 0.000
47
Table VII
Eigenvector Centrality Summary Statistics
The table reports averages of summary statistics for log(eigenvector centrality). To compute averages in the third
and fourth columns, I use customer–supplier data on years 1982, 1987, 1992, 1997, and 2002 to be consistent with the
years used by Ahern (2013). Using data reported in Ahern (2013, Internet Appendix Table II), the second column
presents averages of the statistics for log(eigenvector centrality) in inter-sectoral trade networks. The third column
presents averages in annual customer–supplier networks in which two firms are connected if one firm represents at
least 10% of the other firm’s annual sales. The fourth column presents averages in annual customer–supplier networks
in which two firms are connected if one firm represents at least 20% of the other firm’s annual sales. The fifth column
reports the statistics for log(eigenvector centrality) in the network of the calibrated economy.
Statistic Inter-sectoral Customer–Supplier Customer–Supplier CalibratedNetworks Networks (10%) Networks (20%) Network
Number of sectors/firms 474 750 382 400Mean −6.68 −6.74 −6.62 −6.09Standard deviation 1.48 1.07 1.31 1.71Skewness 0.87 4.04 3.28 1.54Kurtosis 4.45 18.50 12.38 3.70Minimum −10.21 −7.01 −7.01 −7.011st percentile −9.39 −7.01 −7.01 −7.0125th percentile −7.71 −7.01 −7.01 −7.01Median −6.85 −7.01 −7.01 −6.0975th percentile −5.90 −7.01 −7.01 −6.4299th −2.27 −1.83 −1.67 −2.30Maximum −0.17 −0.46 −0.34 −0.74
Table VIII
Performance of Centrality Portfolios
The table reports monthly average raw returns, alphas and loadings from the five-factor model for three portfolios
of stocks sorted by annual centrality and the portfolio that is long on the low tercile and short on the high tercile of
centrality. The bottom row provides the t-statistics for the low minus high portfolio. Firms are assigned into terciles
at the end of October every year and the value-weighted portfolios are not rebalanced for the next 12 months. The
sample is from June 1981 to December 2004. Raw returns and alphas are in percent.
Raw Alphas 5-Factor LoadingsTercile Return CAPM 3-Factor 4-Factor 5-Factor MKT SMB HML RMW CMA
Low 2.06 0.89 1.12 1.01 1.26 0.94 0.12 -0.34 -0.36 0.112 1.96 0.86 0.87 0.83 0.77 1.00 0.06 -0.12 0.07 0.22High 1.71 0.64 0.72 0.75 0.72 0.94 -0.23 -0.11 -0.05 0.05Low - High 0.34 0.25 0.40 0.25 0.54 -0.006 0.36 -0.23 -0.30 0.05t-statistic [1.72] [1.26] [2.3] [1.44] [3.05] [-0.13] [6.10] [-2.67] [-3.94] [0.44]
48
Table IX
Performance of Centrality Portfolios in Manufacturing Stocks
The table reports monthly average raw returns, alphas and loadings from the five-factor model for three portfolios of
manufacturing stocks sorted by annual centrality and the portfolio that is long on the low tercile and short on the
high tercile of centrality. The bottom row provides the t-statistics for the low minus high portfolio. Manufacturing
firms are assigned into terciles at the end of October every year and the value-weighted portfolios are not rebalanced
for the next 12 months. The sample considers from June 1981 to December 2004. Raw returns and alphas are in
percent. Manufacturing stocks represent about 50% of the stocks in the database.
Raw Alphas 5-Factor LoadingsTercile Return CAPM 3-Factor 4-Factor 5-Factor MKT SMB HML RMW CMA
Low 2.02 0.85 1.11 1.01 1.14 0.93 0.29 -0.54 -0.20 0.292 1.85 0.74 0.84 0.87 0.75 0.97 0.13 -0.30 -0.005 0.33High 1.79 0.71 0.82 0.85 0.81 0.93 -0.22 -0.19 -0.09 0.15Low - High 0.23 0.13 0.28 0.15 0.33 0.002 0.52 -0.34 -0.11 0.14t-statistic [1.02] [0.58] [1.49] [0.80] [1.64] [0.04] [7.76] [-3.50] [-1.27] [1.08]
Table X
Performance of Centrality Portfolios in Service Stocks
The table reports monthly average raw returns, alphas and loadings from the five-factor model for three portfolios of
service stocks sorted by annual centrality and the portfolio that is long on the low tercile and short on the high tercile
of centrality. The bottom row provides the t-statistics for the low minus high portfolio. Service firms are assigned
into terciles at the end of October every year and the value-weighted portfolios are not rebalanced for the next 12
months. The sample considers from June 1981 to December 2004. Raw returns and alphas are in percent. Service
stocks represent about 20% of the stocks in the database.
Raw Alphas 5-Factor LoadingsTercile Return CAPM 3-Factor 4-Factor 5-Factor MKT SMB HML RMW CMA
Low 2.86 1.38 1.87 1.69 2.34 1.14 0.58 -0.42 -0.71 -0.462 2.32 1.03 1.34 1.27 1.39 1.07 0.40 -0.43 -0.009 -0.09High 1.54 0.19 0.87 1.13 0.99 1.00 0.06 -0.67 0.15 -0.66Low - High 1.31 1.19 1.00 0.55 1.35 0.14 0.51 0.24 -0.87 0.20t-statistic [2.38] [2.14] [1.79] [0.98] [2.36] [1.03] [2.73] [0.88] [-3.58] [0.54]
Table XI
Performance of Centrality Portfolios in Manufacturing and Service Stocks
The table reports monthly average raw returns, alphas and loadings from the five-factor model for three portfolios of
manufacturing and service stocks sorted by annual centrality and the portfolio that is long on the low tercile and short
on the high tercile of centrality. The bottom row provides the t-statistics for the low-high portfolio. Manufacturing
and service firms are assigned into terciles at the end of October every year and the value-weighted portfolios are
not rebalanced for the next 12 months. The sample considers from June 1981 to December 2004. Raw returns and
alphas are in percent. Manufacturing and service stocks represent about 70% of the stocks in the database.
Raw Alphas 5-Factor LoadingsTercile Return CAPM 3-Factor 4-Factor 5-Factor MKT SMB HML RMW CMA
Low 2.17 0.93 1.26 1.13 1.35 0.98 0.28 -0.59 -0.30 0.242 1.98 0.85 0.99 0.99 0.96 0.96 0.15 -0.33 -0.06 0.25High 1.80 0.71 0.87 0.91 0.92 0.92 -0.24 -0.18 -0.13 -0.01Low - High 0.36 0.22 0.39 0.21 0.42 0.05 0.52 -0.41 -0.17 0.26t-statistic [1.48] [0.92] [1.91] [1.04] [2.00] [1.11] [7.42] [-3.95] [-1.87] [1.89]
49
q
pt
pt
pt
pt
q qq
q
(a) Network Topology
−1 −0.5 0 0.5 1 1.5 2 2.5 30
0.2
0.4
0.6
0.8
1
1.2
1.4
√
nWn
Empirical Probability Density Function of√
nWn
p
t = 0.10
pt = 0.50
pt = 0.90
(b) Empirical density function of√nWn,t
Figure 2. The figure illustrates how changes in the propensity of inter-firm relationships to transmit shocks at t, pt, affect the distributionof
√nWn,t. Figure 2(a) depicts an economy with n = 5 firms, whereas figure 2(b) depicts estimates of the density function of
√nWn,t for
different values of pt. These estimates are computed via normal kernel smoothing estimators using function ksdensity(·) in MATLAB.
50
(a) 2D honeycomb lattice (b) 2D square lattice
(c) 2D triangular lattice (d) Bethe lattice with degree 3
Figure 3. The figure shows the topologies of the symmetric networks considered in corollary 1 and Table I.
51
(a) 1980 (b) 1981 (c) 1982
(d) 1983 (e) 1984 (f) 1985
Figure 4. The figure shows customer–supplier networks from 1980 to 1985.
52
(a) 1986 (b) 1987 (c) 1988
(d) 1989 (e) 1990 (f) 1991
Figure 5. The figure shows customer–supplier networks from 1986 to 1991.
53
(a) 1992 (b) 1993 (c) 1994
(d) 1995 (e) 1996 (f) 1997
Figure 6. The figure shows customer–supplier networks from 1992 to 1997.
54
Histogram of degree in benchmark parametrization
Firms’ degree
Rel
ativ
e fr
eque
ncy
0 10 20 30 40 50
0.0
0.2
0.4
0.6
Figure 7. The figure shows the degree distribution under benchmark parameterization (shown with bars). Dots represent a power lawdistribution with exponent 2.36.
55
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
(a) 1980
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
6070
(b) 1981
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
(c) 1982
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
(d) 1983
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
6070
(e) 1984
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08
010
2030
4050
(f) 1985
Figure 8. The figure shows annual density functions of relationship propensities in customer–supplier networks from 1980 to 1985. Inthe above figures, bars depict annual empirical probability density functions whereas red dots comes from fitting a beta distribution toeach empirical probability density function.
56
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
(a) 1986
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10 0.12
010
2030
4050
60
(b) 1987
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10 0.12
010
2030
4050
60
(c) 1988
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
(d) 1989
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
(e) 1990
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
80
(f) 1991
Figure 9. The figure shows annual density functions of relationship propensities in customer–supplier networks from 1986 to 1991. Inthe above figures, bars depict annual empirical probability density functions whereas red dots comes from fitting a beta distribution toeach empirical probability density function.
57
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
(a) 1992
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
80
(b) 1993
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08
010
2030
4050
6070
(c) 1994
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
(d) 1995
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
(e) 1996
Empirical PDF
Relationships’ propensities
Den
sity
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
(f) 1997
Figure 10. The figure shows annual density functions of relationship propensities in customer–supplier networks from 1992 to 1997. Inthe above figures, bars depict annual empirical probability density functions whereas red dots comes from fitting a beta distribution toeach empirical probability density function.
58
1980 1985 1990 1995 2000
Dynamics of ζ1 and ζ2
Years
ζ 1
0.8
0.9
11.
11.
21.
31.
4
4550
5560
6570
7580
85ζ 2
Figure 11. The figure shows annual estimates of parameters ζ1 and ζ2. I obtain these estimates by fitting Beta distributions to theannual link weight distributions using maximum likelihood.
59
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1980
Corr(Betweenness, Propensity) = −0.25192441667464
(a) 1980
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1981
Corr(Betweenness, Propensity) = −0.246581059581128
(b) 1981
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1982
Corr(Betweenness, Propensity) = −0.288723496355779
(c) 1982
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
6070
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1983
Corr(Betweenness, Propensity) = −0.282322944823283
(d) 1983
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
60
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1984
Corr(Betweenness, Propensity) = −0.200234547101865
(e) 1984
0.00 0.02 0.04 0.06 0.08
020
4060
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1985
Corr(Betweenness, Propensity) = −0.216984491658349
(f) 1985
Figure 12. The figure shows the relationship between link betweenness centrality and link weights in customer–supplier networks from1980 to 1985.
60
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1986
Corr(Betweenness, Propensity) = −0.174316057734155
(a) 1986
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1987
Corr(Betweenness, Propensity) = −0.193516557924177
(b) 1987
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1988
Corr(Betweenness, Propensity) = −0.181563957658669
(c) 1988
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1989
Corr(Betweenness, Propensity) = −0.251258694321243
(d) 1989
0.00 0.02 0.04 0.06 0.08 0.10
010
2030
4050
6070
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1990
Corr(Betweenness, Propensity) = −0.271628985195598
(e) 1990
0.00 0.02 0.04 0.06 0.08
020
4060
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1991
Corr(Betweenness, Propensity) = −0.266126871721907
(f) 1991
Figure 13. The figure shows the relationship between link betweenness centrality and link weights in customer–supplier networks from1986 to 1991.
61
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1992
Corr(Betweenness, Propensity) = −0.190530710570375
(a) 1992
0.00 0.02 0.04 0.06 0.08 0.10
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1993
Corr(Betweenness, Propensity) = −0.214123790383786
(b) 1993
0.00 0.02 0.04 0.06 0.08
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1994
Corr(Betweenness, Propensity) = −0.265006043386911
(c) 1994
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1995
Corr(Betweenness, Propensity) = −0.270367880817289
(d) 1995
0.00 0.02 0.04 0.06 0.08 0.10 0.12
020
4060
80
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1996
Corr(Betweenness, Propensity) = −0.211538705066722
(e) 1996
0.00 0.02 0.04 0.06 0.08 0.10 0.12
010
2030
4050
60
Relationship’s propensity
Rel
atio
nshi
p’s
betw
eenn
ess
Betweenness and Propensity Scores for Relationships in 1997
Corr(Betweenness, Propensity) = −0.1539285869054
(f) 1997
Figure 14. The figure shows the relationship between link betweenness centrality and link weights in customer–supplier networks from1992 to 1997.
62
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1Probabilities pij
0
5
10
15
20
25
30
35
40
Den
sity
func
tion
β(ζLL), β(ζLH), β(ζHL), and β(ζHH)
LLLHHLHH
Figure 15. The figure shows probability density functions for Beta distributions with shape parameter vectors ζLL, ζLH , ζHL, and ζHH .
63
.1.2
.3.4
.5α 0
1980 1985 1990 1995 2000 2005year
Annual estimates of α0
0.0
2.0
4.0
6.0
8.1
α 1
1980 1985 1990 1995 2000 2005year
Annual estimates of α1
Figure 16. The figure shows annual estimates of parameters α0 and α1 in equation (1).
64
0 10 20 30 40 50 60 70 80 900.044
0.045
0.046
0.047
0.048
0.049
0.05Total Volatility by Size Quintile
Years
1 (Small)2345 (Large)
(a) Total Volatility by Size Quintile
0 10 20 30 40 50 60 70 80 90
2.8
3
3.2
3.4
3.6
3.8
4
4.2
4.4x 10
−3 Idiosyncratic Volatility by Size Quintile
Years
1 (Small)2345 (Large)
(b) Idiosyncratic Volatility by Size Quintile
Figure 17. Annualized firm-level volatilities averaged within size quintiles. I simulate 200 panels with 400 firms over 1, 500 periodsand disregard the first 500 periods to eliminate biases coming from the initial condition. Within each panel, I compute firm-level totalvolatility as the annualized standard deviation of monthly firm level realized returns. I construct firm-level idiosyncratic volatility as theannualized standard deviation of residuals computed from monthly CAPM regressions of firm-level excess realized returns on the excessrealized return of the market portfolio. This procedure yields panels of firm-year total and idiosyncratic volatilities estimates. Then,at the beginning of each year I sort firms based on size and average annual volatilities within size quintiles. This procedure yields fivetime series of total and idiosyncratic volatilities per panel. Figure 17(a) plots total volatilities per quintile, whereas figure 17(b) plotsidiosyncratic volatilities per quintile averaged over the 200 panels.
65